/**
* Function for finding and opening either all MCA components, or the one
* that was specifically requested via a MCA parameter.
*/
static int mca_spml_base_open(mca_base_open_flag_t flags)
{
/**
* Construct the send and receive request queues. There are 2 reasons to do it
* here. First, as they are globals it's better to construct them in one common
* place. Second, in order to be able to allow the external debuggers to show
* their content, they should get constructed as soon as possible once the MPI
* process is started.
*/
OBJ_CONSTRUCT(&mca_spml_base_put_requests, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_spml_base_get_requests, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_spml_base_spml, opal_pointer_array_t);
/* Open up all available components */
if (OPAL_SUCCESS !=
mca_base_framework_components_open(&oshmem_spml_base_framework, flags)) {
return OSHMEM_ERROR;
}
/* Set a sentinel in case we don't select any components (e.g.,
ompi_info) */
mca_spml_base_selected_component.spmlm_finalize = NULL;
/**
* Right now our selection of BTLs is completely broken. If we have
* multiple SPMLs that use BTLs than we will open all BTLs several times, leading to
* undefined behaviors. The simplest solution, at least until we
* figure out the correct way to do it, is to force a default SPML that
* uses BTLs and any other SPMLs that do not in the mca_spml_base_spml array.
*/
#if MCA_ompi_pml_DIRECT_CALL
opal_pointer_array_add(&mca_spml_base_spml,
strdup(stringify(MCA_oshmem_spml_DIRECT_CALL_COMPONENT)));
#else
{
const char **default_spml = NULL;
int var_id;
var_id = mca_base_var_find("oshmem", "spml", NULL, NULL);
mca_base_var_get_value(var_id, &default_spml, NULL, NULL);
if( (NULL == default_spml || NULL == default_spml[0] ||
0 == strlen(default_spml[0])) || (default_spml[0][0] == '^') ) {
#ifdef OSHMEM_HAS_IKRIT
opal_pointer_array_add(&mca_spml_base_spml, strdup("ikrit"));
#endif
opal_pointer_array_add(&mca_spml_base_spml, strdup("yoda"));
} else {
opal_pointer_array_add(&mca_spml_base_spml, strdup(default_spml[0]));
}
}
#endif
return OSHMEM_SUCCESS;
}
int orte_snapc_base_select(bool seed, bool app)
{
int exit_status = OPAL_SUCCESS;
orte_snapc_base_component_t *best_component = NULL;
orte_snapc_base_module_t *best_module = NULL;
const char **include_list = NULL;
int var_id;
/*
* Register the framework MCA param and look up include list
*/
/* XXX -- TODO -- framework_subsytem -- this shouldn't be necessary once the framework system is in place */
var_id = mca_base_var_find(NULL, "snapc", NULL, NULL);
mca_base_var_get_value(var_id, &include_list, NULL, NULL);
if(NULL != include_list && NULL != include_list[0] &&
0 == strncmp(include_list[0], "none", strlen("none")) ){
opal_output_verbose(10, orte_snapc_base_framework.framework_output,
"snapc:select: Using %s component",
include_list[0]);
best_module = &none_module;
/* Close all components since none will be used */
mca_base_components_close(0, /* Pass 0 to keep this from closing the output handle */
&orte_snapc_base_framework.framework_components,
NULL);
/* JJH: Todo: Check if none is in the list */
goto skip_select;
}
/*
* Select the best component
*/
if( OPAL_SUCCESS != mca_base_select("snapc", orte_snapc_base_framework.framework_output,
&orte_snapc_base_framework.framework_components,
(mca_base_module_t **) &best_module,
(mca_base_component_t **) &best_component) ) {
/* This will only happen if no component was selected */
exit_status = ORTE_ERROR;
goto cleanup;
}
skip_select:
/* Save the winner */
orte_snapc = *best_module;
/* Initialize the winner */
if (NULL != best_module) {
if (OPAL_SUCCESS != orte_snapc.snapc_init(seed, app)) {
exit_status = OPAL_ERROR;
goto cleanup;
}
}
cleanup:
return exit_status;
}
static int orte_rmaps_rank_file_close(void)
{
int tmp = mca_base_var_find("orte", "orte", NULL, "rankfile");
if (0 <= tmp) {
mca_base_var_deregister(tmp);
}
return ORTE_SUCCESS;
}
int
ompi_init_preconnect_mpi(void)
{
int comm_size = ompi_comm_size(MPI_COMM_WORLD);
int comm_rank = ompi_comm_rank(MPI_COMM_WORLD);
int param, next, prev, i, ret = OMPI_SUCCESS;
struct ompi_request_t * requests[2];
char inbuf[1], outbuf[1];
const bool *value;
param = mca_base_var_find("ompi", "mpi", NULL, "preconnect_mpi");
if (0 > param) return OMPI_SUCCESS;
ret = mca_base_var_get_value(param, &value, NULL, NULL);
if (OMPI_SUCCESS != ret || 0 == value[0]) {
return OMPI_SUCCESS;
}
inbuf[0] = outbuf[0] = '\0';
/* Each iteration, every process sends to its neighbor i hops to
the right and receives from its neighbor i hops to the left.
Because send_complete is used, there will only ever be one
outstanding send and one outstanding receive in the network at
a time for any given process. This limits any "flooding"
effect that can occur with other connection algorithms. While
the flooding algorithms may be a more efficient use of
resources, they can overwhelm the out-of-band connection system
used to wire up some networks, leading to poor performance and
hangs. */
for (i = 1 ; i <= comm_size / 2 ; ++i) {
next = (comm_rank + i) % comm_size;
prev = (comm_rank - i + comm_size) % comm_size;
ret = MCA_PML_CALL(isend(outbuf, 1, MPI_CHAR,
next, 1,
MCA_PML_BASE_SEND_COMPLETE,
MPI_COMM_WORLD,
&requests[1]));
if (OMPI_SUCCESS != ret) return ret;
ret = MCA_PML_CALL(irecv(inbuf, 1, MPI_CHAR,
prev, 1,
MPI_COMM_WORLD,
&requests[0]));
if(OMPI_SUCCESS != ret) return ret;
ret = ompi_request_wait_all(2, requests, MPI_STATUSES_IGNORE);
if (OMPI_SUCCESS != ret) return ret;
}
return ret;
}
int mca_io_ompio_get_fcoll_dynamic_num_io_procs (int *num_procs)
{
int param;
param = mca_base_var_find("ompi", "fcoll", "dynamic", "num_io_procs");
if (param >= 0){
const int *value = NULL;
mca_base_var_get_value(param, &value, NULL, NULL);
*num_procs = value[0];
/* printf("num procs : %d\n", num_procs);*/
return OMPI_SUCCESS;
}
else
return -1;
}
static int tuned_open(void)
{
int rc;
#if OPAL_ENABLE_DEBUG
{
int param;
param = mca_base_var_find("ompi", "coll", "base", "verbose");
if (param >= 0) {
const int *verbose = NULL;
mca_base_var_get_value(param, &verbose, NULL, NULL);
if (verbose && verbose[0] > 0) {
ompi_coll_tuned_stream = opal_output_open(NULL);
}
}
}
#endif /* OPAL_ENABLE_DEBUG */
/* now check that the user hasn't overrode any of the decision functions if dynamic rules are enabled */
/* the user can redo this before every comm dup/create if they like */
/* this is useful for benchmarking and user knows best tuning */
/* as this is the component we only lookup the indicies of the mca params */
/* the actual values are looked up during comm create via module init */
/* intra functions first */
/* if dynamic rules allowed then look up dynamic rules config filename, else we leave it an empty filename (NULL) */
/* by default DISABLE dynamic rules and instead use fixed [if based] rules */
if (ompi_coll_tuned_use_dynamic_rules) {
if( ompi_coll_tuned_dynamic_rules_filename ) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:component_open Reading collective rules file [%s]",
ompi_coll_tuned_dynamic_rules_filename));
rc = ompi_coll_tuned_read_rules_config_file( ompi_coll_tuned_dynamic_rules_filename,
&(mca_coll_tuned_component.all_base_rules), COLLCOUNT);
if( rc >= 0 ) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_open Read %d valid rules\n", rc));
} else {
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:module_open Reading collective rules file failed\n"));
mca_coll_tuned_component.all_base_rules = NULL;
}
}
}
OPAL_OUTPUT((ompi_coll_tuned_stream, "coll:tuned:component_open: done!"));
return OMPI_SUCCESS;
}
int mca_io_ompio_get_fcoll_dynamic_cycle_buffer_size (int *cycle_buffer_size)
{
int param;
param = mca_base_var_find("ompi", "fcoll", "dynamic", "cycle_buffer_size");
if (param >= 0){
const int *value = NULL;
mca_base_var_get_value(param, &value, NULL, NULL);
*cycle_buffer_size = value[0];
/* printf ("cycle_buffer_size : %d\n", *cycle_buffer_size);*/
return OMPI_SUCCESS;
}
else
return -1;
}
int mca_io_ompio_get_fcoll_dynamic_constant_cbs (int *constant_cbs)
{
int param;
param = mca_base_var_find("ompi", "fcoll", "dynamic", "constant_cbs");
if (param >= 0){
const int *value = NULL;
mca_base_var_get_value(param, &value, NULL, NULL);
*constant_cbs = value[0];
/* printf ("constant_cbs: %d\n", constant_cbs);*/
return OMPI_SUCCESS;
}
else{
constant_cbs[0] = -1;
return OMPI_SUCCESS;
}
}
static bool
check_config_value_equal(char *key, ompi_info_t *info, char *value)
{
char *value_string;
int value_len, ret, flag, param;
const bool *flag_value;
bool result = false;
ret = ompi_info_get_valuelen(info, key, &value_len, &flag);
if (OMPI_SUCCESS != ret) goto info_not_found;
if (flag == 0) goto info_not_found;
value_len++;
value_string = (char*)malloc(sizeof(char) * value_len + 1); /* Should malloc 1 char for NUL-termination */
if (NULL == value_string) goto info_not_found;
ret = ompi_info_get(info, key, value_len, value_string, &flag);
if (OMPI_SUCCESS != ret) {
free(value_string);
goto info_not_found;
}
assert(flag != 0);
if (0 == strcmp(value_string, value)) result = true;
free(value_string);
return result;
info_not_found:
param = mca_base_var_find("ompi", "osc", "portals4", key);
if (0 > param) return false;
ret = mca_base_var_get_value(param, &flag_value, NULL, NULL);
if (OMPI_SUCCESS != ret) return false;
if (0 == strcmp(value_string, value)) result = true;
return result;
}
int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
{
int ret;
ompi_proc_t** procs;
size_t nprocs;
char *error = NULL;
struct timeval ompistart, ompistop;
bool rte_setup = false;
ompi_rte_collective_t *coll;
char *cmd=NULL, *av=NULL;
/* bitflag of the thread level support provided. To be used
* for the modex in order to work in heterogeneous environments. */
uint8_t threadlevel_bf;
/* Indicate that we have *started* MPI_INIT*. MPI_FINALIZE has
something sorta similar in a static local variable in
ompi_mpi_finalize(). */
ompi_mpi_init_started = true;
/* Setup enough to check get/set MCA params */
if (OPAL_SUCCESS != (ret = opal_init_util(&argc, &argv))) {
error = "ompi_mpi_init: opal_init_util failed";
goto error;
}
/* Register MCA variables */
if (OPAL_SUCCESS != (ret = ompi_register_mca_variables())) {
error = "ompi_mpi_init: ompi_register_mca_variables failed";
goto error;
}
if (OPAL_SUCCESS != (ret = opal_arch_set_fortran_logical_size(sizeof(ompi_fortran_logical_t)))) {
error = "ompi_mpi_init: opal_arch_set_fortran_logical_size failed";
goto error;
}
/* _After_ opal_init_util() but _before_ orte_init(), we need to
set an MCA param that tells libevent that it's ok to use any
mechanism in libevent that is available on this platform (e.g.,
epoll and friends). Per opal/event/event.s, we default to
select/poll -- but we know that MPI processes won't be using
pty's with the event engine, so it's ok to relax this
constraint and let any fd-monitoring mechanism be used. */
ret = mca_base_var_find("opal", "event", "*", "event_include");
if (ret >= 0) {
char *allvalue = "all";
/* We have to explicitly "set" the MCA param value here
because libevent initialization will re-register the MCA
param and therefore override the default. Setting the value
here puts the desired value ("all") in different storage
that is not overwritten if/when the MCA param is
re-registered. This is unless the user has specified a different
value for this MCA parameter. Make sure we check to see if the
default is specified before forcing "all" in case that is not what
the user desires. Note that we do *NOT* set this value as an
environment variable, just so that it won't be inherited by
any spawned processes and potentially cause unintented
side-effects with launching RTE tools... */
mca_base_var_set_value(ret, allvalue, 4, MCA_BASE_VAR_SOURCE_DEFAULT, NULL);
}
if (ompi_enable_timing && 0 == OMPI_PROC_MY_NAME->vpid) {
gettimeofday(&ompistart, NULL);
}
/* if we were not externally started, then we need to setup
* some envars so the MPI_INFO_ENV can get the cmd name
* and argv (but only if the user supplied a non-NULL argv!), and
* the requested thread level
*/
if (NULL == getenv("OMPI_COMMAND") && NULL != argv && NULL != argv[0]) {
asprintf(&cmd, "OMPI_COMMAND=%s", argv[0]);
putenv(cmd);
}
if (NULL == getenv("OMPI_ARGV") && 1 < argc) {
char *tmp;
tmp = opal_argv_join(&argv[1], ' ');
asprintf(&av, "OMPI_ARGV=%s", tmp);
free(tmp);
putenv(av);
}
/* open the rte framework */
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_rte_base_framework, 0))) {
error = "ompi_rte_base_open() failed";
goto error;
}
/* no select is required as this is a static framework */
/* Setup RTE - note that we are an MPI process */
if (OMPI_SUCCESS != (ret = ompi_rte_init(NULL, NULL))) {
error = "ompi_mpi_init: ompi_rte_init failed";
goto error;
}
rte_setup = true;
/* check for timing request - get stop time and report elapsed time if so */
if (ompi_enable_timing && 0 == OMPI_PROC_MY_NAME->vpid) {
gettimeofday(&ompistop, NULL);
opal_output(0, "ompi_mpi_init [%ld]: time from start to completion of rte_init %ld usec",
(long)OMPI_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
gettimeofday(&ompistart, NULL);
}
#if OPAL_HAVE_HWLOC
/* if hwloc is available but didn't get setup for some
* reason, do so now
*/
if (NULL == opal_hwloc_topology) {
if (OPAL_SUCCESS != (ret = opal_hwloc_base_get_topology())) {
error = "Topology init";
goto error;
}
}
#endif
/* Register the default errhandler callback - RTE will ignore if it
* doesn't support this capability
*/
ompi_rte_register_errhandler(ompi_errhandler_runtime_callback,
OMPI_RTE_ERRHANDLER_LAST);
/* Figure out the final MPI thread levels. If we were not
compiled for support for MPI threads, then don't allow
MPI_THREAD_MULTIPLE. Set this stuff up here early in the
process so that other components can make decisions based on
this value. */
ompi_mpi_thread_level(requested, provided);
/* determine the bitflag belonging to the threadlevel_support provided */
memset ( &threadlevel_bf, 0, sizeof(uint8_t));
OMPI_THREADLEVEL_SET_BITFLAG ( ompi_mpi_thread_provided, threadlevel_bf );
/* add this bitflag to the modex */
if ( OMPI_SUCCESS != (ret = ompi_modex_send_string("MPI_THREAD_LEVEL", &threadlevel_bf, sizeof(uint8_t)))) {
error = "ompi_mpi_init: modex send thread level";
goto error;
}
/* initialize datatypes. This step should be done early as it will
* create the local convertor and local arch used in the proc
* init.
*/
if (OMPI_SUCCESS != (ret = ompi_datatype_init())) {
error = "ompi_datatype_init() failed";
goto error;
}
/* Initialize OMPI procs */
if (OMPI_SUCCESS != (ret = ompi_proc_init())) {
error = "mca_proc_init() failed";
goto error;
}
/* Initialize the op framework. This has to be done *after*
ddt_init, but befor mca_coll_base_open, since some collective
modules (e.g., the hierarchical coll component) may need ops in
their query function. */
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_op_base_framework, 0))) {
error = "ompi_op_base_open() failed";
goto error;
}
if (OMPI_SUCCESS !=
(ret = ompi_op_base_find_available(OMPI_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_THREAD_MULTIPLE))) {
error = "ompi_op_base_find_available() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = ompi_op_init())) {
error = "ompi_op_init() failed";
goto error;
}
/* Open up MPI-related MCA components */
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_allocator_base_framework, 0))) {
error = "mca_allocator_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_rcache_base_framework, 0))) {
error = "mca_rcache_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_mpool_base_framework, 0))) {
error = "mca_mpool_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_bml_base_framework, 0))) {
error = "mca_bml_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_pml_base_framework, 0))) {
error = "mca_pml_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_coll_base_framework, 0))) {
error = "mca_coll_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_osc_base_framework, 0))) {
error = "ompi_osc_base_open() failed";
goto error;
}
#if OPAL_ENABLE_FT_CR == 1
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_crcp_base_framework, 0))) {
error = "ompi_crcp_base_open() failed";
goto error;
}
#endif
/* In order to reduce the common case for MPI apps (where they
don't use MPI-2 IO or MPI-1 topology functions), the io and
topo frameworks are initialized lazily, at the first use of
relevant functions (e.g., MPI_FILE_*, MPI_CART_*, MPI_GRAPH_*),
so they are not opened here. */
/* Select which MPI components to use */
if (OMPI_SUCCESS !=
(ret = mca_mpool_base_init(OMPI_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_THREAD_MULTIPLE))) {
error = "mca_mpool_base_init() failed";
goto error;
}
if (OMPI_SUCCESS !=
(ret = mca_pml_base_select(OMPI_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_THREAD_MULTIPLE))) {
error = "mca_pml_base_select() failed";
goto error;
}
/* check for timing request - get stop time and report elapsed time if so */
if (ompi_enable_timing && 0 == OMPI_PROC_MY_NAME->vpid) {
gettimeofday(&ompistop, NULL);
opal_output(0, "ompi_mpi_init[%ld]: time from completion of rte_init to modex %ld usec",
(long)OMPI_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
gettimeofday(&ompistart, NULL);
}
/* exchange connection info - this function also acts as a barrier
* as it will not return until the exchange is complete
*/
coll = OBJ_NEW(ompi_rte_collective_t);
coll->id = ompi_process_info.peer_modex;
coll->active = true;
if (OMPI_SUCCESS != (ret = ompi_rte_modex(coll))) {
error = "rte_modex failed";
goto error;
}
/* wait for modex to complete - this may be moved anywhere in mpi_init
* so long as it occurs prior to calling a function that needs
* the modex info!
*/
OMPI_WAIT_FOR_COMPLETION(coll->active);
OBJ_RELEASE(coll);
if (ompi_enable_timing && 0 == OMPI_PROC_MY_NAME->vpid) {
gettimeofday(&ompistop, NULL);
opal_output(0, "ompi_mpi_init[%ld]: time to execute modex %ld usec",
(long)OMPI_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
gettimeofday(&ompistart, NULL);
}
/* select buffered send allocator component to be used */
if( OMPI_SUCCESS !=
(ret = mca_pml_base_bsend_init(OMPI_ENABLE_THREAD_MULTIPLE))) {
error = "mca_pml_base_bsend_init() failed";
goto error;
}
if (OMPI_SUCCESS !=
(ret = mca_coll_base_find_available(OMPI_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_THREAD_MULTIPLE))) {
error = "mca_coll_base_find_available() failed";
goto error;
}
if (OMPI_SUCCESS !=
(ret = ompi_osc_base_find_available(OMPI_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_THREAD_MULTIPLE))) {
error = "ompi_osc_base_find_available() failed";
goto error;
}
#if OPAL_ENABLE_FT_CR == 1
if (OMPI_SUCCESS != (ret = ompi_crcp_base_select() ) ) {
error = "ompi_crcp_base_select() failed";
goto error;
}
#endif
/* io and topo components are not selected here -- see comment
above about the io and topo frameworks being loaded lazily */
/* Initialize each MPI handle subsystem */
/* initialize requests */
if (OMPI_SUCCESS != (ret = ompi_request_init())) {
error = "ompi_request_init() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = ompi_message_init())) {
error = "ompi_message_init() failed";
goto error;
}
/* initialize info */
if (OMPI_SUCCESS != (ret = ompi_info_init())) {
error = "ompi_info_init() failed";
goto error;
}
/* initialize error handlers */
if (OMPI_SUCCESS != (ret = ompi_errhandler_init())) {
error = "ompi_errhandler_init() failed";
goto error;
}
/* initialize error codes */
if (OMPI_SUCCESS != (ret = ompi_mpi_errcode_init())) {
error = "ompi_mpi_errcode_init() failed";
goto error;
}
/* initialize internal error codes */
if (OMPI_SUCCESS != (ret = ompi_errcode_intern_init())) {
error = "ompi_errcode_intern_init() failed";
goto error;
}
/* initialize groups */
if (OMPI_SUCCESS != (ret = ompi_group_init())) {
error = "ompi_group_init() failed";
goto error;
}
/* initialize communicators */
if (OMPI_SUCCESS != (ret = ompi_comm_init())) {
error = "ompi_comm_init() failed";
goto error;
}
/* initialize file handles */
if (OMPI_SUCCESS != (ret = ompi_file_init())) {
error = "ompi_file_init() failed";
goto error;
}
/* initialize windows */
if (OMPI_SUCCESS != (ret = ompi_win_init())) {
error = "ompi_win_init() failed";
goto error;
}
/* initialize attribute meta-data structure for comm/win/dtype */
if (OMPI_SUCCESS != (ret = ompi_attr_init())) {
error = "ompi_attr_init() failed";
goto error;
}
/* identify the architectures of remote procs and setup
* their datatype convertors, if required
*/
if (OMPI_SUCCESS != (ret = ompi_proc_complete_init())) {
error = "ompi_proc_complete_init failed";
goto error;
}
/* If thread support was enabled, then setup OPAL to allow for
them. */
if ((OMPI_ENABLE_PROGRESS_THREADS == 1) ||
(*provided != MPI_THREAD_SINGLE)) {
opal_set_using_threads(true);
}
/* start PML/BTL's */
ret = MCA_PML_CALL(enable(true));
if( OMPI_SUCCESS != ret ) {
error = "PML control failed";
goto error;
}
/* add all ompi_proc_t's to PML */
if (NULL == (procs = ompi_proc_world(&nprocs))) {
error = "ompi_proc_world() failed";
goto error;
}
ret = MCA_PML_CALL(add_procs(procs, nprocs));
free(procs);
/* If we got "unreachable", then print a specific error message.
Otherwise, if we got some other failure, fall through to print
a generic message. */
if (OMPI_ERR_UNREACH == ret) {
opal_show_help("help-mpi-runtime",
"mpi_init:startup:pml-add-procs-fail", true);
error = NULL;
goto error;
} else if (OMPI_SUCCESS != ret) {
error = "PML add procs failed";
goto error;
}
MCA_PML_CALL(add_comm(&ompi_mpi_comm_world.comm));
MCA_PML_CALL(add_comm(&ompi_mpi_comm_self.comm));
/*
* Dump all MCA parameters if requested
*/
if (ompi_mpi_show_mca_params) {
ompi_show_all_mca_params(ompi_mpi_comm_world.comm.c_my_rank,
nprocs,
ompi_process_info.nodename);
}
/* Do we need to wait for a debugger? */
ompi_rte_wait_for_debugger();
/* check for timing request - get stop time and report elapsed
time if so, then start the clock again */
if (ompi_enable_timing && 0 == OMPI_PROC_MY_NAME->vpid) {
gettimeofday(&ompistop, NULL);
opal_output(0, "ompi_mpi_init[%ld]: time from modex to first barrier %ld usec",
(long)OMPI_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
gettimeofday(&ompistart, NULL);
}
/* wait for everyone to reach this point */
coll = OBJ_NEW(ompi_rte_collective_t);
coll->id = ompi_process_info.peer_init_barrier;
coll->active = true;
if (OMPI_SUCCESS != (ret = ompi_rte_barrier(coll))) {
error = "rte_barrier failed";
goto error;
}
/* wait for barrier to complete */
OMPI_WAIT_FOR_COMPLETION(coll->active);
OBJ_RELEASE(coll);
/* check for timing request - get stop time and report elapsed
time if so, then start the clock again */
if (ompi_enable_timing && 0 == OMPI_PROC_MY_NAME->vpid) {
gettimeofday(&ompistop, NULL);
opal_output(0, "ompi_mpi_init[%ld]: time to execute barrier %ld usec",
(long)OMPI_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
gettimeofday(&ompistart, NULL);
}
#if OMPI_ENABLE_PROGRESS_THREADS == 0
/* Start setting up the event engine for MPI operations. Don't
block in the event library, so that communications don't take
forever between procs in the dynamic code. This will increase
CPU utilization for the remainder of MPI_INIT when we are
blocking on RTE-level events, but may greatly reduce non-TCP
latency. */
opal_progress_set_event_flag(OPAL_EVLOOP_NONBLOCK);
#endif
/* wire up the mpi interface, if requested. Do this after the
non-block switch for non-TCP performance. Do before the
polling change as anyone with a complex wire-up is going to be
using the oob. */
if (OMPI_SUCCESS != (ret = ompi_init_preconnect_mpi())) {
error = "ompi_mpi_do_preconnect_all() failed";
goto error;
}
/* Setup the publish/subscribe (PUBSUB) framework */
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_pubsub_base_framework, 0))) {
error = "mca_pubsub_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = ompi_pubsub_base_select())) {
error = "ompi_pubsub_base_select() failed";
goto error;
}
/* Setup the dynamic process management (DPM) framework */
if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_dpm_base_framework, 0))) {
error = "ompi_dpm_base_open() failed";
goto error;
}
if (OMPI_SUCCESS != (ret = ompi_dpm_base_select())) {
error = "ompi_dpm_base_select() failed";
goto error;
}
/* Determine the overall threadlevel support of all processes
in MPI_COMM_WORLD. This has to be done before calling
coll_base_comm_select, since some of the collective components
e.g. hierarch, might create subcommunicators. The threadlevel
requested by all processes is required in order to know
which cid allocation algorithm can be used. */
if ( OMPI_SUCCESS !=
( ret = ompi_comm_cid_init ())) {
error = "ompi_mpi_init: ompi_comm_cid_init failed";
goto error;
}
/* Init coll for the comms. This has to be after dpm_base_select,
(since dpm.mark_dyncomm is not set in the communicator creation
function else), but before dpm.dyncom_init, since this function
might require collective for the CID allocation. */
if (OMPI_SUCCESS !=
(ret = mca_coll_base_comm_select(MPI_COMM_WORLD))) {
error = "mca_coll_base_comm_select(MPI_COMM_WORLD) failed";
goto error;
}
if (OMPI_SUCCESS !=
(ret = mca_coll_base_comm_select(MPI_COMM_SELF))) {
error = "mca_coll_base_comm_select(MPI_COMM_SELF) failed";
goto error;
}
/* Check whether we have been spawned or not. We introduce that
at the very end, since we need collectives, datatypes, ptls
etc. up and running here.... */
if (OMPI_SUCCESS != (ret = ompi_dpm.dyn_init())) {
error = "ompi_comm_dyn_init() failed";
goto error;
}
/*
* Startup the Checkpoint/Restart Mech.
* Note: Always do this so tools don't hang when
* in a non-checkpointable build
*/
if (OMPI_SUCCESS != (ret = ompi_cr_init())) {
error = "ompi_cr_init";
goto error;
}
/* Undo OPAL calling opal_progress_event_users_increment() during
opal_init, to get better latency when not using TCP. Do
this *after* dyn_init, as dyn init uses lots of RTE
communication and we don't want to hinder the performance of
that code. */
opal_progress_event_users_decrement();
/* see if yield_when_idle was specified - if so, use it */
opal_progress_set_yield_when_idle(ompi_mpi_yield_when_idle);
/* negative value means use default - just don't do anything */
if (ompi_mpi_event_tick_rate >= 0) {
opal_progress_set_event_poll_rate(ompi_mpi_event_tick_rate);
}
/* At this point, we are fully configured and in MPI mode. Any
communication calls here will work exactly like they would in
the user's code. Setup the connections between procs and warm
them up with simple sends, if requested */
if (OMPI_SUCCESS != (ret = ompi_mpiext_init())) {
error = "ompi_mpiext_init";
goto error;
}
/* Fall through */
error:
if (ret != OMPI_SUCCESS) {
/* Only print a message if one was not already printed */
if (NULL != error) {
const char *err_msg = opal_strerror(ret);
/* If RTE was not setup yet, don't use opal_show_help */
if (rte_setup) {
opal_show_help("help-mpi-runtime",
"mpi_init:startup:internal-failure", true,
"MPI_INIT", "MPI_INIT", error, err_msg, ret);
} else {
opal_show_help("help-mpi-runtime",
"mpi_init:startup:internal-failure", true,
"MPI_INIT", "MPI_INIT", error, err_msg, ret);
}
}
return ret;
}
/* Initialize the registered datarep list to be empty */
OBJ_CONSTRUCT(&ompi_registered_datareps, opal_list_t);
/* Initialize the arrays used to store the F90 types returned by the
* MPI_Type_create_f90_XXX functions.
*/
OBJ_CONSTRUCT( &ompi_mpi_f90_integer_hashtable, opal_hash_table_t);
opal_hash_table_init(&ompi_mpi_f90_integer_hashtable, 16 /* why not? */);
OBJ_CONSTRUCT( &ompi_mpi_f90_real_hashtable, opal_hash_table_t);
opal_hash_table_init(&ompi_mpi_f90_real_hashtable, FLT_MAX_10_EXP);
OBJ_CONSTRUCT( &ompi_mpi_f90_complex_hashtable, opal_hash_table_t);
opal_hash_table_init(&ompi_mpi_f90_complex_hashtable, FLT_MAX_10_EXP);
/* All done. Wasn't that simple? */
ompi_mpi_initialized = true;
/* check for timing request - get stop time and report elapsed time if so */
if (ompi_enable_timing && 0 == OMPI_PROC_MY_NAME->vpid) {
gettimeofday(&ompistop, NULL);
opal_output(0, "ompi_mpi_init[%ld]: time from barrier to complete mpi_init %ld usec",
(long)OMPI_PROC_MY_NAME->vpid,
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
(ompistop.tv_usec - ompistart.tv_usec)));
}
return MPI_SUCCESS;
}
int mca_bml_r2_ft_event(int state)
{
#if OPAL_ENABLE_FT_CR == 1
static bool first_continue_pass = false;
ompi_proc_t** procs = NULL;
size_t num_procs;
size_t btl_idx;
int ret, p;
int loc_state;
int param_type = -1;
const char **btl_list;
if(OPAL_CRS_CHECKPOINT == state) {
/* Do nothing for now */
}
else if(OPAL_CRS_CONTINUE == state) {
first_continue_pass = !first_continue_pass;
/* Since nothing in Checkpoint, we are fine here (unless required by BTL) */
if (opal_cr_continue_like_restart && !first_continue_pass) {
procs = ompi_proc_all(&num_procs);
if(NULL == procs) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
}
else if(OPAL_CRS_RESTART_PRE == state ) {
/* Nothing here */
}
else if(OPAL_CRS_RESTART == state ) {
procs = ompi_proc_all(&num_procs);
if(NULL == procs) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
else if(OPAL_CRS_TERM == state ) {
;
}
else {
;
}
/* Never call the ft_event functions attached to the BTLs on the second
* pass of RESTART since on the first pass they were unloaded and therefore
* no longer exist.
*/
if( OPAL_CRS_RESTART != state ) {
if( OPAL_CRS_CONTINUE == state && !first_continue_pass ) {
;
} else {
/* Since we only ever call into the BTLs once during the first restart
* pass, just lie to them on this pass for a bit of local clarity.
*/
if( OPAL_CRS_RESTART_PRE == state ) {
loc_state = OPAL_CRS_RESTART;
} else {
loc_state = state;
}
/*
* Call ft_event in:
* - BTL modules
* - MPool modules
*
* These should be cleaning out stale state, and memory references in
* preparation for being shut down.
*/
for(btl_idx = 0; btl_idx < mca_bml_r2.num_btl_modules; btl_idx++) {
/*
* Notify Mpool
*/
if( NULL != (mca_bml_r2.btl_modules[btl_idx])->btl_mpool &&
NULL != (mca_bml_r2.btl_modules[btl_idx])->btl_mpool->mpool_ft_event ) {
opal_output_verbose(10, ompi_cr_output,
"bml:r2: ft_event: Notify the %s MPool.\n",
(mca_bml_r2.btl_modules[btl_idx])->btl_mpool->mpool_component->mpool_version.mca_component_name);
if(OMPI_SUCCESS != (ret = (mca_bml_r2.btl_modules[btl_idx])->btl_mpool->mpool_ft_event(loc_state) ) ) {
continue;
}
}
/*
* Notify BTL
*/
if( NULL != (mca_bml_r2.btl_modules[btl_idx])->btl_ft_event) {
opal_output_verbose(10, ompi_cr_output,
"bml:r2: ft_event: Notify the %s BTL.\n",
(mca_bml_r2.btl_modules[btl_idx])->btl_component->btl_version.mca_component_name);
if(OMPI_SUCCESS != (ret = (mca_bml_r2.btl_modules[btl_idx])->btl_ft_event(loc_state) ) ) {
continue;
}
}
}
} /* OPAL_CRS_CONTINUE == state && !first_continue_pass */
}
if(OPAL_CRS_CHECKPOINT == state) {
;
}
else if(OPAL_CRS_CONTINUE == state) {
/* Matches OPAL_CRS_RESTART_PRE */
if (opal_cr_continue_like_restart && first_continue_pass) {
if( OMPI_SUCCESS != (ret = mca_bml_r2_finalize()) ) {
opal_output(0, "bml:r2: ft_event(Restart): Failed to finalize BML framework\n");
return ret;
}
if( OMPI_SUCCESS != (ret = mca_base_framework_close(&opal_btl_base_framework)) ) {
opal_output(0, "bml:r2: ft_event(Restart): Failed to close BTL framework\n");
return ret;
}
}
/* Matches OPAL_CRS_RESTART */
else if (opal_cr_continue_like_restart && !first_continue_pass) {
/*
* Barrier to make all processes have been successfully restarted before
* we try to remove some restart only files.
*/
opal_pmix.fence(NULL, 0);
/*
* Re-open the BTL framework to get the full list of components.
*/
if( OMPI_SUCCESS != (ret = mca_base_framework_open(&opal_btl_base_framework, 0)) ) {
opal_output(0, "bml:r2: ft_event(Restart): Failed to open BTL framework\n");
return ret;
}
/*
* Re-select the BTL components/modules
* This will cause the BTL components to discover the available
* network options on this machine, and post proper modex informaiton.
*/
if( OMPI_SUCCESS != (ret = mca_btl_base_select(OPAL_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_THREAD_MULTIPLE) ) ) {
opal_output(0, "bml:r2: ft_event(Restart): Failed to select in BTL framework\n");
return ret;
}
/*
* Clear some structures so we can properly repopulate them
*/
mca_bml_r2.btls_added = false;
for(p = 0; p < (int)num_procs; ++p) {
if( NULL != procs[p]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]) {
OBJ_RELEASE(procs[p]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]);
procs[p]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML] = NULL;
}
OBJ_RELEASE(procs[p]);
}
if( NULL != procs ) {
free(procs);
procs = NULL;
}
}
}
else if(OPAL_CRS_RESTART_PRE == state ) {
opal_output_verbose(10, ompi_cr_output,
"bml:r2: ft_event(Restart): Finalize BML\n");
/*
* Finalize the BML
* - Flush progress functions
* - Flush module references
* - mca_btl_base_close()
* Need to do this because we may have BTL components that were
* unloaded in the first selection that may be available now.
* Conversely we may have BTL components loaded now that
* are not available now.
*/
if( OMPI_SUCCESS != (ret = mca_bml_r2_finalize()) ) {
opal_output(0, "bml:r2: ft_event(Restart): Failed to finalize BML framework\n");
return ret;
}
if( OMPI_SUCCESS != (ret = mca_base_framework_close(&opal_btl_base_framework)) ) {
opal_output(0, "bml:r2: ft_event(Restart): Failed to close BTL framework\n");
return ret;
}
}
else if(OPAL_CRS_RESTART == state ) {
/*
* Barrier to make all processes have been successfully restarted before
* we try to remove some restart only files.
*/
opal_pmix.fence(NULL, 0);
/*
* Re-open the BTL framework to get the full list of components.
* - but first clear the MCA value that was there
*/
param_type = mca_base_var_find("ompi", "btl", NULL, NULL);
btl_list = NULL;
mca_base_var_get_value(param_type, &btl_list, NULL, NULL);
opal_output_verbose(11, ompi_cr_output,
"Restart (Previous BTL MCA): <%s>\n", btl_list ? btl_list[0] : "");
if( OMPI_SUCCESS != (ret = mca_base_framework_open(&opal_btl_base_framework, 0)) ) {
opal_output(0, "bml:r2: ft_event(Restart): Failed to open BTL framework\n");
return ret;
}
/* The reregistered paramter is guaranteed to have the same index */
btl_list = NULL;
mca_base_var_get_value(param_type, &btl_list, NULL, NULL);
opal_output_verbose(11, ompi_cr_output,
"Restart (New BTL MCA): <%s>\n", btl_list ? btl_list[0] : "");
if( NULL != btl_list ) {
free(btl_list);
btl_list = NULL;
}
/*
* Re-select the BTL components/modules
* This will cause the BTL components to discover the available
* network options on this machine, and post proper modex informaiton.
*/
if( OMPI_SUCCESS != (ret = mca_btl_base_select(OPAL_ENABLE_PROGRESS_THREADS,
OMPI_ENABLE_THREAD_MULTIPLE) ) ) {
opal_output(0, "bml:r2: ft_event(Restart): Failed to select in BTL framework\n");
return ret;
}
/*
* Clear some structures so we can properly repopulate them
*/
mca_bml_r2.btls_added = false;
for(p = 0; p < (int)num_procs; ++p) {
if( NULL != procs[p]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]) {
OBJ_RELEASE(procs[p]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]);
procs[p]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML] = NULL;
}
OBJ_RELEASE(procs[p]);
}
if( NULL != procs ) {
free(procs);
procs = NULL;
}
}
else if(OPAL_CRS_TERM == state ) {
;
}
else {
;
}
#endif
return OMPI_SUCCESS;
}
int
main(int argc, char *argv[])
{
int ret, exit_status = OPAL_SUCCESS;
int child_pid;
int prev_pid = 0;
int idx;
opal_crs_base_snapshot_t *snapshot = NULL;
char * tmp_env_var = NULL;
bool select = false;
/***************
* Initialize
***************/
if (OPAL_SUCCESS != (ret = initialize(argc, argv))) {
exit_status = ret;
goto cleanup;
}
/*
* Check for existence of the file, or program in the case of self
*/
if( OPAL_SUCCESS != (ret = check_file() )) {
opal_show_help("help-opal-restart.txt", "invalid_filename", true,
opal_restart_globals.snapshot_ref);
exit_status = ret;
goto cleanup;
}
/* Re-enable the selection of the CRS component, so we can choose the right one */
idx = mca_base_var_find(NULL, "crs", "base", "do_not_select");
if (0 > idx) {
opal_output(opal_restart_globals.output,
"MCA variable opal_crs_base_do_not_select not found\n");
exit_status = OPAL_ERROR;
goto cleanup;
}
ret = mca_base_var_set_value(idx, &select, 0, MCA_BASE_VAR_SOURCE_DEFAULT, NULL);
if (OPAL_SUCCESS != ret) {
exit_status = ret;
goto cleanup;
}
/*
* Make sure we are using the correct checkpointer
*/
if(NULL == expected_crs_comp) {
char * full_metadata_path = NULL;
FILE * metadata = NULL;
opal_asprintf(&full_metadata_path, "%s/%s/%s",
opal_restart_globals.snapshot_loc,
opal_restart_globals.snapshot_ref,
opal_restart_globals.snapshot_metadata);
if( NULL == (metadata = fopen(full_metadata_path, "r")) ) {
opal_show_help("help-opal-restart.txt", "invalid_metadata", true,
opal_restart_globals.snapshot_metadata,
full_metadata_path);
exit_status = OPAL_ERROR;
goto cleanup;
}
if( OPAL_SUCCESS != (ret = opal_crs_base_extract_expected_component(metadata,
&expected_crs_comp,
&prev_pid)) ) {
opal_show_help("help-opal-restart.txt", "invalid_metadata", true,
opal_restart_globals.snapshot_metadata,
full_metadata_path);
exit_status = ret;
goto cleanup;
}
free(full_metadata_path);
full_metadata_path = NULL;
fclose(metadata);
metadata = NULL;
}
opal_output_verbose(10, opal_restart_globals.output,
"Restart Expects checkpointer: (%s)",
expected_crs_comp);
(void) mca_base_var_env_name("crs", &tmp_env_var);
opal_setenv(tmp_env_var,
expected_crs_comp,
true, &environ);
free(tmp_env_var);
tmp_env_var = NULL;
/* Select this component or don't continue.
* If the selection of this component fails, then we can't
* restart on this node because it doesn't have the proper checkpointer
* available.
*/
if( OPAL_SUCCESS != (ret = opal_crs_base_open(MCA_BASE_OPEN_DEFAULT)) ) {
opal_show_help("help-opal-restart.txt", "comp_select_failure", true,
"crs", ret);
exit_status = ret;
goto cleanup;
}
if( OPAL_SUCCESS != (ret = opal_crs_base_select()) ) {
opal_show_help("help-opal-restart.txt", "comp_select_failure", true,
expected_crs_comp, ret);
exit_status = ret;
goto cleanup;
}
/*
* Make sure we have selected the proper component
*/
if(NULL == expected_crs_comp ||
0 != strncmp(expected_crs_comp,
opal_crs_base_selected_component.base_version.mca_component_name,
strlen(expected_crs_comp)) ) {
opal_show_help("help-opal-restart.txt", "comp_select_mismatch",
true,
expected_crs_comp,
opal_crs_base_selected_component.base_version.mca_component_name,
ret);
exit_status = ret;
goto cleanup;
}
/******************************
* Restart in this process
******************************/
opal_output_verbose(10, opal_restart_globals.output,
"Restarting from file (%s)\n",
opal_restart_globals.snapshot_ref);
snapshot = OBJ_NEW(opal_crs_base_snapshot_t);
snapshot->cold_start = true;
opal_asprintf(&(snapshot->snapshot_directory), "%s/%s",
opal_restart_globals.snapshot_loc,
opal_restart_globals.snapshot_ref);
opal_asprintf(&(snapshot->metadata_filename), "%s/%s",
snapshot->snapshot_directory,
opal_restart_globals.snapshot_metadata);
/* Since some checkpoint/restart systems don't pass along env vars to the
* restarted app, we need to take care of that.
*
* Included here is the creation of any files or directories that need to be
* created before the process is restarted.
*/
if(OPAL_SUCCESS != (ret = post_env_vars(prev_pid, snapshot) ) ) {
exit_status = ret;
goto cleanup;
}
/*
* Do the actual restart
*/
ret = opal_crs.crs_restart(snapshot,
false,
&child_pid);
if (OPAL_SUCCESS != ret) {
opal_show_help("help-opal-restart.txt", "restart_cmd_failure", true,
opal_restart_globals.snapshot_ref,
ret,
opal_crs_base_selected_component.base_version.mca_component_name);
exit_status = ret;
goto cleanup;
}
/* Should never get here, since crs_restart calls exec */
/***************
* Cleanup
***************/
cleanup:
if (OPAL_SUCCESS != (ret = finalize())) {
return ret;
}
if(NULL != snapshot )
OBJ_DESTRUCT(snapshot);
return exit_status;
}
int ompi_mpi_register_params(void)
{
int value;
/* Whether we want MPI API function parameter checking or not. Disable this by default if
parameter checking is compiled out. */
ompi_mpi_param_check = !!(MPI_PARAM_CHECK);
(void) mca_base_var_register("ompi", "mpi", NULL, "param_check",
"Whether you want MPI API parameters checked at run-time or not. Possible values are 0 (no checking) and 1 (perform checking at run-time)",
MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mpi_param_check);
if (ompi_mpi_param_check && !MPI_PARAM_CHECK) {
opal_show_help("help-mpi-runtime.txt",
"mpi-param-check-enabled-but-compiled-out",
true);
ompi_mpi_param_check = false;
}
/*
* opal_progress: decide whether to yield and the event library
* tick rate
*/
/* JMS: Need ORTE data here -- set this to 0 when
exactly/under-subscribed, or 1 when oversubscribed */
ompi_mpi_yield_when_idle = true;
(void) mca_base_var_register("ompi", "mpi", NULL, "yield_when_idle",
"Yield the processor when waiting for MPI communication (for MPI processes, will default to 1 when oversubscribing nodes)",
MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mpi_yield_when_idle);
ompi_mpi_event_tick_rate = -1;
(void) mca_base_var_register("ompi", "mpi", NULL, "event_tick_rate",
"How often to progress TCP communications (0 = never, otherwise specified in microseconds)",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mpi_event_tick_rate);
/* Whether or not to show MPI handle leaks */
ompi_debug_show_handle_leaks = false;
(void) mca_base_var_register("ompi", "mpi", NULL, "show_handle_leaks",
"Whether MPI_FINALIZE shows all MPI handles that were not freed or not",
MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_debug_show_handle_leaks);
/* Whether or not to free MPI handles. Useless without run-time
param checking, so implicitly set that to true if we don't want
to free the handles. */
ompi_debug_no_free_handles = false;
(void) mca_base_var_register("ompi", "mpi", NULL, "no_free_handles",
"Whether to actually free MPI objects when their handles are freed",
MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_debug_no_free_handles);
if (ompi_debug_no_free_handles) {
ompi_mpi_param_check = true;
if (!MPI_PARAM_CHECK) {
opal_output(0, "WARNING: MCA parameter mpi_no_free_handles set to true, but MPI");
opal_output(0, "WARNING: parameter checking has been compiled out of Open MPI.");
opal_output(0, "WARNING: mpi_no_free_handles is therefore only partially effective!");
}
}
/* Whether or not to show MPI_ALLOC_MEM leaks */
ompi_debug_show_mpi_alloc_mem_leaks = 0;
(void) mca_base_var_register("ompi", "mpi", NULL, "show_mpi_alloc_mem_leaks",
"If >0, MPI_FINALIZE will show up to this many instances of memory allocated by MPI_ALLOC_MEM that was not freed by MPI_FREE_MEM",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_debug_show_mpi_alloc_mem_leaks);
/* Whether or not to print all MCA parameters in MPI_INIT */
ompi_mpi_show_mca_params_string = NULL;
(void) mca_base_var_register("ompi", "mpi", NULL, "show_mca_params",
"Whether to show all MCA parameter values during MPI_INIT or not (good for reproducability of MPI jobs "
"for debug purposes). Accepted values are all, default, file, api, and enviro - or a comma "
"delimited combination of them",
MCA_BASE_VAR_TYPE_STRING, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mpi_show_mca_params_string);
if (NULL != ompi_mpi_show_mca_params_string) {
char **args;
int i;
ompi_mpi_show_mca_params = true;
args = opal_argv_split(ompi_mpi_show_mca_params_string, ',');
if (NULL == args) {
opal_output(0, "WARNING: could not parse mpi_show_mca_params request - defaulting to show \"all\"");
show_default_mca_params = true;
show_file_mca_params = true;
show_enviro_mca_params = true;
show_override_mca_params = true;
} else {
for (i=0; NULL != args[i]; i++) {
if (0 == strcasecmp(args[i], "all") || 0 == strcmp(args[i], "1")) {
show_default_mca_params = true;
show_file_mca_params = true;
show_enviro_mca_params = true;
show_override_mca_params = true;
} else if (0 == strcasecmp(args[i], "default")) {
show_default_mca_params = true;
} else if (0 == strcasecmp(args[i], "file")) {
show_file_mca_params = true;
} else if (0 == strncasecmp(args[i], "env", 3)) {
show_enviro_mca_params = true;
} else if (0 == strcasecmp(args[i], "api")) {
show_override_mca_params = true;
}
}
opal_argv_free(args);
}
}
/* File to use when dumping the parameters */
ompi_mpi_show_mca_params_file = "";
(void) mca_base_var_register("ompi", "mpi", NULL, "show_mca_params_file",
"If mpi_show_mca_params is true, setting this string to a valid filename tells Open MPI to dump all the MCA parameter values into a file suitable for reading via the mca_param_files parameter (good for reproducability of MPI jobs)",
MCA_BASE_VAR_TYPE_STRING, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mpi_show_mca_params_file);
/* User-level process pinning controls */
/* MPI_ABORT controls */
ompi_mpi_abort_delay = 0;
(void) mca_base_var_register("ompi", "mpi", NULL, "abort_delay",
"If nonzero, print out an identifying message when MPI_ABORT is invoked (hostname, PID of the process that called MPI_ABORT) and delay for that many seconds before exiting (a negative delay value means to never abort). This allows attaching of a debugger before quitting the job.",
MCA_BASE_VAR_TYPE_INT, NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mpi_abort_delay);
ompi_mpi_abort_print_stack = false;
(void) mca_base_var_register("ompi", "mpi", NULL, "abort_print_stack",
"If nonzero, print out a stack trace when MPI_ABORT is invoked",
MCA_BASE_VAR_TYPE_BOOL, NULL, 0,
/* If we do not have stack trace
capability, make this a constant
MCA variable */
#if OPAL_WANT_PRETTY_PRINT_STACKTRACE
0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
#else
MCA_BASE_VAR_FLAG_DEFAULT_ONLY,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_CONSTANT,
#endif
&ompi_mpi_abort_print_stack);
ompi_mpi_preconnect_mpi = false;
value = mca_base_var_register("ompi", "mpi", NULL, "preconnect_mpi",
"Whether to force MPI processes to fully "
"wire-up the MPI connections between MPI "
"processes during "
"MPI_INIT (vs. making connections lazily -- "
"upon the first MPI traffic between each "
"process peer pair)",
MCA_BASE_VAR_TYPE_BOOL, NULL, 0,
MCA_BASE_VAR_FLAG_INTERNAL,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&ompi_mpi_preconnect_mpi);
mca_base_var_register_synonym(value, "ompi", "mpi", NULL, "preconnect_all",
MCA_BASE_VAR_SYN_FLAG_DEPRECATED);
/* Sparse group storage support */
(void) mca_base_var_register("ompi", "mpi", NULL, "have_sparse_group_storage",
"Whether this Open MPI installation supports storing of data in MPI groups in \"sparse\" formats (good for extremely large process count MPI jobs that create many communicators/groups)",
MCA_BASE_VAR_TYPE_BOOL, NULL, 0,
MCA_BASE_VAR_FLAG_DEFAULT_ONLY,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_CONSTANT,
&ompi_mpi_have_sparse_group_storage);
ompi_use_sparse_group_storage = ompi_mpi_have_sparse_group_storage;
(void) mca_base_var_register("ompi", "mpi", NULL, "use_sparse_group_storage",
"Whether to use \"sparse\" storage formats for MPI groups (only relevant if mpi_have_sparse_group_storage is 1)",
MCA_BASE_VAR_TYPE_BOOL, NULL, 0,
ompi_mpi_have_sparse_group_storage ? 0 : MCA_BASE_VAR_FLAG_DEFAULT_ONLY,
OPAL_INFO_LVL_9,
ompi_mpi_have_sparse_group_storage ? MCA_BASE_VAR_SCOPE_READONLY : MCA_BASE_VAR_SCOPE_CONSTANT,
&ompi_use_sparse_group_storage);
if (ompi_use_sparse_group_storage && !ompi_mpi_have_sparse_group_storage) {
opal_show_help("help-mpi-runtime.txt",
"sparse groups enabled but compiled out",
true);
ompi_use_sparse_group_storage = false;
}
value = mca_base_var_find ("opal", "opal", NULL, "cuda_support");
if (0 <= value) {
mca_base_var_register_synonym(value, "ompi", "mpi", NULL, "cuda_support",
MCA_BASE_VAR_SYN_FLAG_DEPRECATED);
}
value = mca_base_var_find ("opal", "opal", NULL, "built_with_cuda_support");
if (0 <= value) {
mca_base_var_register_synonym(value, "ompi", "mpi", NULL, "built_with_cuda_support", 0);
}
if (opal_cuda_support && !opal_built_with_cuda_support) {
opal_show_help("help-mpi-runtime.txt", "no cuda support",
true);
ompi_rte_abort(1, NULL);
}
ompi_add_procs_cutoff = 1024;
(void) mca_base_var_register ("ompi", "mpi", NULL, "add_procs_cutoff",
"Maximum world size for pre-allocating resources for all "
"remote processes. Increasing this limit may improve "
"communication performance at the cost of memory usage "
"(default: 1024)", MCA_BASE_VAR_TYPE_UNSIGNED_INT, NULL,
0, 0, OPAL_INFO_LVL_3, MCA_BASE_VAR_SCOPE_LOCAL,
&ompi_add_procs_cutoff);
return OMPI_SUCCESS;
}
RTE_PUBLIC int rte_orte_init(int *argc, char ***argv, rte_group_t *out_group)
{
int rc;
if (rte_initialized)
goto exit;
rte_initialized = true;
if (OPAL_SUCCESS != (rc = opal_init_util(NULL, NULL))) {
return RTE_ERROR;
}
/* Pasha: This is MPI specific code
if (OPAL_SUCCESS != (ret = opal_arch_set_fortran_logical_size(sizeof(ompi_fortran_logical_t)))) {
return RTE_ERROR;
}
*/
/* _After_ opal_init_util() but _before_ orte_init(), we need to
set an MCA param that tells libevent that it's ok to use any
mechanism in libevent that is available on this platform (e.g.,
epoll and friends). Per opal/event/event.s, we default to
select/poll -- but we know that MPI processes won't be using
pty's with the event engine, so it's ok to relax this
constraint and let any fd-monitoring mechanism be used. */
rc = mca_base_var_find("opal", "event", "*", "event_include");
if (rc >= 0) {
char *allvalue = "all";
/* We have to explicitly "set" the MCA param value here
because libevent initialization will re-register the MCA
param and therefore override the default. Setting the value
here puts the desired value ("all") in different storage
that is not overwritten if/when the MCA param is
re-registered. This is unless the user has specified a different
value for this MCA parameter. Make sure we check to see if the
default is specified before forcing "all" in case that is not what
the user desires. Note that we do *NOT* set this value as an
environment variable, just so that it won't be inherited by
any spawned processes and potentially cause unintented
side-effects with launching RTE tools... */
mca_base_var_set_value(rc, allvalue, 4, MCA_BASE_VAR_SOURCE_DEFAULT, NULL);
}
opal_set_using_threads(false); /* true - just doesn't work */
/* Pasha: Let's present ourself as ompi proc */
rc = orte_init(NULL, NULL, ORTE_PROC_MPI);
if (OPAL_SUCCESS != rc) {
*out_group = NULL;
return RTE_ERROR;
}
/* Temporary disabled, since it conflicts with OMPI and we don't really use
it with ddt so far.
hook_debugger();
*/
rc = all_orte_proc_init();
if (RTE_SUCCESS != rc) {
return RTE_ERROR;
}
/* see if yield_when_idle was specified - if so, use it */
opal_progress_set_yield_when_idle (true);
exit:
/* Pasha: Orte has no support for groups, instead we return job object */
*out_group = &orte_process_info;
return RTE_SUCCESS;
error:
return RTE_ERROR;
}
