1
1
openmpi/opal/runtime/opal_progress.c
Ralph Castain 7bee71aa59 Fix a potential, albeit perhaps esoteric, race condition that can occur for fast HNP's, slow orteds, and fast apps. Under those conditions, it is possible for the orted to be caught in its original send of contact info back to the HNP, and thus for the progress stack never to recover back to a high level. In those circumstances, the orted can "hang" when trying to exit.
Add a new function to opal_progress that tells us our recursion depth to support that solution.

Yes, I know this sounds picky, but good ol' Jeff managed to make it happen by driving his cluster near to death...

Also ensure that we declare "failed" for the daemon job when daemons fail instead of the application job. This is important so that orte knows that it cannot use xcast to tell daemons to "exit", nor should it expect all daemons to respond. Otherwise, it is possible to hang.

After lots of testing, decide to default (again) to slurm detecting failed orteds. This proved necessary to avoid rather annoying hangs that were difficult to recover from. There are conditions where slurm will fail to launch all daemons (slurm folks are working on it), and yet again, good ol' Jeff managed to find both of them.

Thanks you Jeff! :-/

This commit was SVN r18611.
2008-06-06 19:36:27 +00:00

410 строки
12 KiB
C

/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2005 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006 Los Alamos National Security, LLC. All rights
* reserved.
*
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "opal_config.h"
#ifdef HAVE_SCHED_H
#include <sched.h>
#endif
#include "opal/runtime/opal_progress.h"
#include "opal/event/event.h"
#include "opal/mca/base/mca_base_param.h"
#include "opal/constants.h"
#include "opal/mca/timer/base/base.h"
#include "opal/util/output.h"
#define OPAL_PROGRESS_USE_TIMERS (OPAL_TIMER_CYCLE_SUPPORTED || OPAL_TIMER_USEC_SUPPORTED)
/*
* default parameters
*/
static int opal_progress_event_flag = OPAL_EVLOOP_ONELOOP;
volatile int32_t opal_progress_thread_count = 0;
int opal_progress_spin_count = 10000;
/*
* Local variables
*/
#if OMPI_HAVE_THREAD_SUPPORT
static opal_atomic_lock_t progress_lock;
#endif /* OMPI_HAVE_THREAD_SUPPORT */
/* callbacks to progress */
static opal_progress_callback_t *callbacks = NULL;
static size_t callbacks_len = 0;
static size_t callbacks_size = 0;
/* do we want to call sched_yield() if nothing happened */
static int call_yield = 1;
#if OPAL_PROGRESS_USE_TIMERS
static opal_timer_t event_progress_last_time = 0;
static opal_timer_t event_progress_delta = 0;
#else
/* current count down until we tick the event library */
static int32_t event_progress_counter = 0;
/* reset value for counter when it hits 0 */
static int32_t event_progress_delta = 0;
#endif
/* users of the event library from MPI cause the tick rate to
be every time */
static int32_t num_event_users = 0;
/* How deep are we in opal_progress recursion? */
#if OMPI_HAVE_THREAD_SUPPORT
volatile
#endif
uint32_t opal_progress_recursion_depth_counter = 0;
#if OMPI_ENABLE_DEBUG
static int debug_output = -1;
#endif
/**
* Fake callback used for threading purpose when one thread
* progesses callbacks while another unregister somes. The root
* of the problem is that we allow modifications of the callback
* array directly from the callbacks themselves. Now if
* writing a pointer is atomic, we should not have any more
* problems.
*/
static int fake_cb(void) { return 0; };
/* init the progress engine - called from orte_init */
int
opal_progress_init(void)
{
#if OMPI_ENABLE_DEBUG
int param, value;
#endif
/* reentrant issues */
#if OMPI_HAVE_THREAD_SUPPORT
opal_atomic_init(&progress_lock, OPAL_ATOMIC_UNLOCKED);
#endif /* OMPI_HAVE_THREAD_SUPPORT */
/* set the event tick rate */
opal_progress_set_event_poll_rate(10000);
#if OMPI_ENABLE_DEBUG
param = mca_base_param_find("opal", NULL, "progress_debug");
mca_base_param_lookup_int(param, &value);
if (value) {
debug_output = opal_output_open(NULL);
}
#endif
OPAL_OUTPUT((debug_output, "progress: initialized event flag to: %x",
opal_progress_event_flag));
OPAL_OUTPUT((debug_output, "progress: initialized yield_when_idle to: %s",
call_yield == 0 ? "false" : "true"));
OPAL_OUTPUT((debug_output, "progress: initialized num users to: %d",
num_event_users));
OPAL_OUTPUT((debug_output, "progress: initialized poll rate to: %ld",
(long) event_progress_delta));
return OPAL_SUCCESS;
}
int
opal_progress_finalize(void)
{
/* free memory associated with the callbacks */
#if OMPI_HAVE_THREAD_SUPPORT
opal_atomic_lock(&progress_lock);
#endif
callbacks_len = 0;
callbacks_size = 0;
if (NULL != callbacks) {
free(callbacks);
callbacks = NULL;
}
#if OMPI_HAVE_THREAD_SUPPORT
opal_atomic_unlock(&progress_lock);
#endif
return OPAL_SUCCESS;
}
/*
* Progress the event library and any functions that have registered to
* be called. We don't propogate errors from the progress functions,
* so no action is taken if they return failures. The functions are
* expected to return the number of events progressed, to determine
* whether or not we should call sched_yield() during MPI progress.
* This is only losely tracked, as an error return can cause the number
* of progressed events to appear lower than it actually is. We don't
* care, as the cost of that happening is far outweighed by the cost
* of the if checks (they were resulting in bad pipe stalling behavior)
*/
void
opal_progress(void)
{
size_t i;
int events = 0;
#if OMPI_HAVE_THREAD_SUPPORT
opal_atomic_add(&opal_progress_recursion_depth_counter, 1);
#else
++opal_progress_recursion_depth_counter;
#endif
if( opal_progress_event_flag != 0 ) {
#if (OMPI_ENABLE_PROGRESS_THREADS == 0) && OPAL_HAVE_WORKING_EVENTOPS
#if OPAL_PROGRESS_USE_TIMERS
#if OPAL_TIMER_USEC_NATIVE
opal_timer_t now = opal_timer_base_get_usec();
#else
opal_timer_t now = opal_timer_base_get_cycles();
#endif /* OPAL_TIMER_USEC_NATIVE */
/* trip the event library if we've reached our tick rate and we are
enabled */
if (now - event_progress_last_time > event_progress_delta ) {
event_progress_last_time = (num_event_users > 0) ?
now - event_progress_delta : now;
events += opal_event_loop(opal_progress_event_flag);
}
#else /* OPAL_PROGRESS_USE_TIMERS */
/* trip the event library if we've reached our tick rate and we are
enabled */
if (OPAL_THREAD_ADD32(&event_progress_counter, -1) <= 0 ) {
event_progress_counter =
(num_event_users > 0) ? 0 : event_progress_delta;
events += opal_event_loop(opal_progress_event_flag);
}
#endif /* OPAL_PROGRESS_USE_TIMERS */
#endif /* OMPI_ENABLE_PROGRESS_THREADS == 0 && OPAL_HAVE_WORKING_EVENTOPS */
}
/* progress all registered callbacks */
for (i = 0 ; i < callbacks_len ; ++i) {
events += (callbacks[i])();
}
#if defined(__WINDOWS__) || defined(HAVE_SCHED_YIELD)
if (call_yield && events <= 0) {
/* If there is nothing to do - yield the processor - otherwise
* we could consume the processor for the entire time slice. If
* the processor is oversubscribed - this will result in a best-case
* latency equivalent to the time-slice.
*/
#if defined(__WINDOWS__)
SwitchToThread();
#else
sched_yield();
#endif /* defined(__WINDOWS__) */
}
#endif /* defined(__WINDOWS__) || defined(HAVE_SCHED_YIELD) */
#if OMPI_HAVE_THREAD_SUPPORT
opal_atomic_add(&opal_progress_recursion_depth_counter, -1);
#else
--opal_progress_recursion_depth_counter;
#endif
}
int
opal_progress_set_event_flag(int flag)
{
int tmp = opal_progress_event_flag;
opal_progress_event_flag = flag;
OPAL_OUTPUT((debug_output, "progress: set_event_flag setting to %d", flag));
return tmp;
}
void
opal_progress_event_users_increment(void)
{
int32_t val;
val = opal_atomic_add_32(&num_event_users, 1);
OPAL_OUTPUT((debug_output, "progress: event_users_increment setting count to %d", val));
#if OPAL_PROGRESS_USE_TIMERS
/* force an update next round (we'll be past the delta) */
event_progress_last_time -= event_progress_delta;
#else
/* always reset the tick rate - can't hurt */
event_progress_counter = 0;
#endif
}
void
opal_progress_event_users_decrement(void)
{
int32_t val;
val = opal_atomic_sub_32(&num_event_users, 1);
OPAL_OUTPUT((debug_output, "progress: event_users_decrement setting count to %d", val));
#if !OPAL_PROGRESS_USE_TIMERS
/* start now in delaying if it's easy */
if (val >= 0) {
event_progress_counter = event_progress_delta;
}
#endif
}
bool
opal_progress_set_yield_when_idle(bool yieldopt)
{
bool tmp = (call_yield == 0) ? false : true;
call_yield = (yieldopt) ? 1 : 0;
OPAL_OUTPUT((debug_output, "progress: progress_set_yield_when_idle to %s",
call_yield == 0 ? "false" : "true"));
return tmp;
}
void
opal_progress_set_event_poll_rate(int polltime)
{
OPAL_OUTPUT((debug_output, "progress: progress_set_event_poll_rate(%d)", polltime));
#if OPAL_PROGRESS_USE_TIMERS
event_progress_delta = 0;
# if OPAL_TIMER_USEC_NATIVE
event_progress_last_time = opal_timer_base_get_usec();
# else
event_progress_last_time = opal_timer_base_get_cycles();
# endif
#else
event_progress_counter = event_progress_delta = 0;
#endif
if (polltime == 0) {
#if OPAL_PROGRESS_USE_TIMERS
/* user specified as never tick - tick once per minute */
event_progress_delta = 60 * 1000000;
#else
/* user specified as never tick - don't count often */
event_progress_delta = INT_MAX;
#endif
} else {
#if OPAL_PROGRESS_USE_TIMERS
event_progress_delta = polltime;
#else
/* subtract one so that we can do post-fix subtraction
in the inner loop and go faster */
event_progress_delta = polltime - 1;
#endif
}
#if OPAL_PROGRESS_USE_TIMERS && !OPAL_TIMER_USEC_NATIVE
/* going to use cycles for counter. Adjust specified usec into cycles */
event_progress_delta = event_progress_delta * opal_timer_base_get_freq() / 1000000;
#endif
}
int
opal_progress_register(opal_progress_callback_t cb)
{
int ret = OPAL_SUCCESS, index;
#if OMPI_HAVE_THREAD_SUPPORT
opal_atomic_lock(&progress_lock);
#endif
/* see if we need to allocate more space */
if (callbacks_len + 1 > callbacks_size) {
opal_progress_callback_t *tmp;
tmp = (opal_progress_callback_t*)realloc(callbacks, sizeof(opal_progress_callback_t) * (callbacks_size + 4));
if (tmp == NULL) {
ret = OPAL_ERR_TEMP_OUT_OF_RESOURCE;
goto cleanup;
}
/* registering fake callbacks to fill callbacks[] */
for( index = callbacks_len + 1 ; index < callbacks_size + 4 ; index++) {
tmp[index] = &fake_cb;
}
callbacks = tmp;
callbacks_size += 4;
}
callbacks[callbacks_len++] = cb;
cleanup:
#if OMPI_HAVE_THREAD_SUPPORT
opal_atomic_unlock(&progress_lock);
#endif
return ret;
}
int
opal_progress_unregister(opal_progress_callback_t cb)
{
size_t i;
int ret = OPAL_ERR_NOT_FOUND;
#if OMPI_HAVE_THREAD_SUPPORT
opal_atomic_lock(&progress_lock);
#endif
for (i = 0 ; i < callbacks_len ; ++i) {
if (cb == callbacks[i]) {
callbacks[i] = &fake_cb;
ret = OPAL_SUCCESS;
break;
}
}
/* If we found the function we're unregistering: If callbacks_len
is 0, we're not goig to do anything interesting anyway, so
skip. If callbacks_len is 1, it will soon be 0, so no need to
do any repacking. size_t can be unsigned, so 0 - 1 is bad for
a loop condition :). */
if (OPAL_SUCCESS == ret) {
if (callbacks_len > 1 ) {
/* now tightly pack the array */
for ( ; i < callbacks_len - 1 ; ++i) {
callbacks[i] = callbacks[i + 1];
}
}
callbacks[callbacks_len - 1] = &fake_cb;
callbacks_len--;
}
#if OMPI_HAVE_THREAD_SUPPORT
opal_atomic_unlock(&progress_lock);
#endif
return ret;
}