a200e4f865
*** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
415 строки
14 KiB
C
415 строки
14 KiB
C
/* -*- Mode: C; c-basic-offset:4 ; -*- */
|
|
/*
|
|
* Copyright (c) 2004-2010 The Trustees of Indiana University and Indiana
|
|
* University Research and Technology
|
|
* Corporation. All rights reserved.
|
|
* Copyright (c) 2004-2011 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) 2006-2012 Cisco Systems, Inc. All rights reserved.
|
|
* Copyright (c) 2006-2012 Los Alamos National Security, LLC. All rights
|
|
* reserved.
|
|
* Copyright (c) 2006 University of Houston. All rights reserved.
|
|
* Copyright (c) 2009 Sun Microsystems, Inc. All rights reserved.
|
|
* Copyright (c) 2011 Sandia National Laboratories. All rights reserved.
|
|
*
|
|
* $COPYRIGHT$
|
|
*
|
|
* Additional copyrights may follow
|
|
*
|
|
* $HEADER$
|
|
*/
|
|
|
|
#include "ompi_config.h"
|
|
|
|
#ifdef HAVE_SYS_TYPES_H
|
|
#include <sys/types.h>
|
|
#endif
|
|
#ifdef HAVE_UNISTD_H
|
|
#include <unistd.h>
|
|
#endif
|
|
#ifdef HAVE_SYS_PARAM_H
|
|
#include <sys/param.h>
|
|
#endif
|
|
#ifdef HAVE_NETDB_H
|
|
#include <netdb.h>
|
|
#endif
|
|
|
|
#include "opal/mca/event/event.h"
|
|
#include "opal/util/output.h"
|
|
#include "opal/runtime/opal_progress.h"
|
|
#include "opal/mca/base/base.h"
|
|
#include "opal/sys/atomic.h"
|
|
#include "opal/runtime/opal.h"
|
|
#include "opal/util/show_help.h"
|
|
|
|
#include "mpi.h"
|
|
#include "ompi/constants.h"
|
|
#include "ompi/errhandler/errcode.h"
|
|
#include "ompi/communicator/communicator.h"
|
|
#include "ompi/datatype/ompi_datatype.h"
|
|
#include "ompi/message/message.h"
|
|
#include "ompi/op/op.h"
|
|
#include "ompi/file/file.h"
|
|
#include "ompi/info/info.h"
|
|
#include "ompi/runtime/mpiruntime.h"
|
|
#include "ompi/attribute/attribute.h"
|
|
#include "ompi/mca/pml/pml.h"
|
|
#include "ompi/mca/pml/base/base.h"
|
|
#include "ompi/mca/osc/base/base.h"
|
|
#include "ompi/mca/coll/base/base.h"
|
|
#include "ompi/mca/rte/rte.h"
|
|
#include "ompi/mca/topo/base/base.h"
|
|
#include "ompi/mca/io/io.h"
|
|
#include "ompi/mca/io/base/base.h"
|
|
#include "ompi/mca/mpool/base/base.h"
|
|
#include "ompi/mca/mpool/base/mpool_base_tree.h"
|
|
#include "ompi/mca/rcache/base/base.h"
|
|
#include "ompi/mca/pml/base/pml_base_bsend.h"
|
|
#include "ompi/runtime/params.h"
|
|
#include "ompi/mca/dpm/base/base.h"
|
|
#include "ompi/mca/pubsub/base/base.h"
|
|
#include "ompi/mpiext/mpiext.h"
|
|
|
|
#if OPAL_ENABLE_FT_CR == 1
|
|
#include "ompi/mca/crcp/crcp.h"
|
|
#include "ompi/mca/crcp/base/base.h"
|
|
#endif
|
|
#include "ompi/runtime/ompi_cr.h"
|
|
|
|
extern bool ompi_enable_timing;
|
|
|
|
int ompi_mpi_finalize(void)
|
|
{
|
|
int ret;
|
|
static int32_t finalize_has_already_started = 0;
|
|
opal_list_item_t *item;
|
|
struct timeval ompistart, ompistop;
|
|
ompi_rte_collective_t *coll;
|
|
|
|
/* Be a bit social if an erroneous program calls MPI_FINALIZE in
|
|
two different threads, otherwise we may deadlock in
|
|
ompi_comm_free() (or run into other nasty lions, tigers, or
|
|
bears) */
|
|
|
|
if (! opal_atomic_cmpset_32(&finalize_has_already_started, 0, 1)) {
|
|
/* Note that if we're already finalized, we cannot raise an
|
|
MPI exception. The best that we can do is write something
|
|
to stderr. */
|
|
char hostname[MAXHOSTNAMELEN];
|
|
pid_t pid = getpid();
|
|
gethostname(hostname, sizeof(hostname));
|
|
|
|
opal_show_help("help-mpi-runtime.txt",
|
|
"mpi_finalize:invoked_multiple_times",
|
|
true, hostname, pid);
|
|
return MPI_ERR_OTHER;
|
|
}
|
|
|
|
ompi_mpiext_fini();
|
|
|
|
/* As finalize is the last legal MPI call, we are allowed to force the release
|
|
* of the user buffer used for bsend, before going anywhere further.
|
|
*/
|
|
(void)mca_pml_base_bsend_detach(NULL, NULL);
|
|
|
|
/* Per MPI-2:4.8, we have to free MPI_COMM_SELF before doing
|
|
anything else in MPI_FINALIZE (to include setting up such that
|
|
MPI_FINALIZED will return true). */
|
|
|
|
if (NULL != ompi_mpi_comm_self.comm.c_keyhash) {
|
|
ompi_attr_delete_all(COMM_ATTR, &ompi_mpi_comm_self,
|
|
ompi_mpi_comm_self.comm.c_keyhash);
|
|
OBJ_RELEASE(ompi_mpi_comm_self.comm.c_keyhash);
|
|
ompi_mpi_comm_self.comm.c_keyhash = NULL;
|
|
}
|
|
|
|
/* Proceed with MPI_FINALIZE */
|
|
|
|
ompi_mpi_finalized = true;
|
|
|
|
#if OMPI_ENABLE_PROGRESS_THREADS == 0
|
|
opal_progress_set_event_flag(OPAL_EVLOOP_ONCE | OPAL_EVLOOP_NONBLOCK);
|
|
#endif
|
|
|
|
/* Redo ORTE calling opal_progress_event_users_increment() during
|
|
MPI lifetime, to get better latency when not using TCP */
|
|
opal_progress_event_users_increment();
|
|
|
|
/* check to see if we want timing information */
|
|
if (ompi_enable_timing != 0 && 0 == OMPI_PROC_MY_NAME->vpid) {
|
|
gettimeofday(&ompistart, NULL);
|
|
}
|
|
|
|
/* NOTE: MPI-2.1 requires that MPI_FINALIZE is "collective" across
|
|
*all* connected processes. This only means that all processes
|
|
have to call it. It does *not* mean that all connected
|
|
processes need to synchronize (either directly or indirectly).
|
|
|
|
For example, it is quite easy to construct complicated
|
|
scenarios where one job is "connected" to another job via
|
|
transitivity, but have no direct knowledge of each other.
|
|
Consider the following case: job A spawns job B, and job B
|
|
later spawns job C. A "connectedness" graph looks something
|
|
like this:
|
|
|
|
A <--> B <--> C
|
|
|
|
So what are we *supposed* to do in this case? If job A is
|
|
still connected to B when it calls FINALIZE, should it block
|
|
until jobs B and C also call FINALIZE?
|
|
|
|
After lengthy discussions many times over the course of this
|
|
project, the issue was finally decided at the Louisville Feb
|
|
2009 meeting: no.
|
|
|
|
Rationale:
|
|
|
|
- "Collective" does not mean synchronizing. It only means that
|
|
every process call it. Hence, in this scenario, every
|
|
process in A, B, and C must call FINALIZE.
|
|
|
|
- KEY POINT: if A calls FINALIZE, then it is erroneous for B or
|
|
C to try to communicate with A again.
|
|
|
|
- Hence, OMPI is *correct* to only effect a barrier across each
|
|
jobs' MPI_COMM_WORLD before exiting. Specifically, if A
|
|
calls FINALIZE long before B or C, it's *correct* if A exits
|
|
at any time (and doesn't notify B or C that it is exiting).
|
|
|
|
- Arguably, if B or C do try to communicate with the now-gone
|
|
A, OMPI should try to print a nice error ("you tried to
|
|
communicate with a job that is already gone...") instead of
|
|
segv or other Badness. However, that is an *extremely*
|
|
difficult problem -- sure, it's easy for A to tell B that it
|
|
is finalizing, but how can A tell C? A doesn't even know
|
|
about C. You'd need to construct a "connected" graph in a
|
|
distributed fashion, which is fraught with race conditions,
|
|
etc.
|
|
|
|
Hence, our conclusion is: OMPI is *correct* in its current
|
|
behavior (of only doing a barrier across its own COMM_WORLD)
|
|
before exiting. Any problems that occur are as a result of
|
|
erroneous MPI applications. We *could* tighten up the erroneous
|
|
cases and ensure that we print nice error messages / don't
|
|
crash, but that is such a difficult problem that we decided we
|
|
have many other, much higher priority issues to handle that deal
|
|
with non-erroneous cases. */
|
|
|
|
/* wait for everyone to reach this point
|
|
This is a grpcomm barrier instead of an MPI barrier because an
|
|
MPI barrier doesn't ensure that all messages have been transmitted
|
|
before exiting, so the possibility of a stranded message exists.
|
|
*/
|
|
coll = OBJ_NEW(ompi_rte_collective_t);
|
|
coll->id = ompi_process_info.peer_fini_barrier;
|
|
coll->active = true;
|
|
if (OMPI_SUCCESS != (ret = ompi_rte_barrier(coll))) {
|
|
OMPI_ERROR_LOG(ret);
|
|
return ret;
|
|
}
|
|
|
|
/* 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 */
|
|
if (ompi_enable_timing && 0 == OMPI_PROC_MY_NAME->vpid) {
|
|
gettimeofday(&ompistop, NULL);
|
|
opal_output(0, "ompi_mpi_finalize[%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)));
|
|
}
|
|
|
|
/*
|
|
* Shutdown the Checkpoint/Restart Mech.
|
|
*/
|
|
if (OMPI_SUCCESS != (ret = ompi_cr_finalize())) {
|
|
OMPI_ERROR_LOG(ret);
|
|
}
|
|
|
|
/* Shut down any bindings-specific issues: C++, F77, F90 */
|
|
|
|
/* Remove all memory associated by MPI_REGISTER_DATAREP (per
|
|
MPI-2:9.5.3, there is no way for an MPI application to
|
|
*un*register datareps, but we don't want the OMPI layer causing
|
|
memory leaks). */
|
|
while (NULL != (item = opal_list_remove_first(&ompi_registered_datareps))) {
|
|
OBJ_RELEASE(item);
|
|
}
|
|
OBJ_DESTRUCT(&ompi_registered_datareps);
|
|
|
|
/* Remove all F90 types from the hash tables. As the OBJ_DESTRUCT will
|
|
* call a special destructor able to release predefined types, we can
|
|
* simply call the OBJ_DESTRUCT on the hash table and all memory will
|
|
* be correctly released.
|
|
*/
|
|
OBJ_DESTRUCT( &ompi_mpi_f90_integer_hashtable );
|
|
OBJ_DESTRUCT( &ompi_mpi_f90_real_hashtable );
|
|
OBJ_DESTRUCT( &ompi_mpi_f90_complex_hashtable );
|
|
|
|
/* Free communication objects */
|
|
|
|
/* free file resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_file_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* free window resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_win_finalize())) {
|
|
return ret;
|
|
}
|
|
if (OMPI_SUCCESS != (ret = ompi_osc_base_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* free pml resource */
|
|
if(OMPI_SUCCESS != (ret = mca_pml_base_finalize())) {
|
|
return ret;
|
|
}
|
|
/* free communicator resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_comm_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* free requests */
|
|
if (OMPI_SUCCESS != (ret = ompi_request_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
if (OMPI_SUCCESS != (ret = ompi_message_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* If requested, print out a list of memory allocated by ALLOC_MEM
|
|
but not freed by FREE_MEM */
|
|
if (0 != ompi_debug_show_mpi_alloc_mem_leaks) {
|
|
mca_mpool_base_tree_print();
|
|
}
|
|
|
|
/* Now that all MPI objects dealing with communications are gone,
|
|
shut down MCA types having to do with communications */
|
|
if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_pml_base_framework) ) ) {
|
|
OMPI_ERROR_LOG(ret);
|
|
return ret;
|
|
}
|
|
|
|
/* shut down buffered send code */
|
|
mca_pml_base_bsend_fini();
|
|
|
|
#if OPAL_ENABLE_FT_CR == 1
|
|
/*
|
|
* Shutdown the CRCP Framework, must happen after PML shutdown
|
|
*/
|
|
if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_crcp_base_framework) ) ) {
|
|
OMPI_ERROR_LOG(ret);
|
|
return ret;
|
|
}
|
|
#endif
|
|
|
|
/* Free secondary resources */
|
|
|
|
/* free attr resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_attr_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* free group resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_group_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* free proc resources */
|
|
if ( OMPI_SUCCESS != (ret = ompi_proc_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* finalize the pubsub functions */
|
|
if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_pubsub_base_framework) ) ) {
|
|
return ret;
|
|
}
|
|
|
|
/* finalize the DPM framework */
|
|
if ( OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_dpm_base_framework))) {
|
|
return ret;
|
|
}
|
|
|
|
/* free internal error resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_errcode_intern_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* free error code resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_mpi_errcode_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* free errhandler resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_errhandler_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* Free all other resources */
|
|
|
|
/* free op resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_op_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* free ddt resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_datatype_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* free info resources */
|
|
if (OMPI_SUCCESS != (ret = ompi_info_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
/* Close down MCA modules */
|
|
|
|
/* io is opened lazily, so it's only necessary to close it if it
|
|
was actually opened */
|
|
if (0 < ompi_io_base_framework.framework_refcnt) {
|
|
/* May have been "opened" multiple times. We want it closed now */
|
|
ompi_io_base_framework.framework_refcnt = 1;
|
|
|
|
if (OMPI_SUCCESS != mca_base_framework_close(&ompi_io_base_framework)) {
|
|
return ret;
|
|
}
|
|
}
|
|
(void) mca_base_framework_close(&ompi_topo_base_framework);
|
|
if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_osc_base_framework))) {
|
|
return ret;
|
|
}
|
|
if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_coll_base_framework))) {
|
|
return ret;
|
|
}
|
|
if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_mpool_base_framework))) {
|
|
return ret;
|
|
}
|
|
if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_rcache_base_framework))) {
|
|
return ret;
|
|
}
|
|
|
|
/* Leave the RTE */
|
|
|
|
if (OMPI_SUCCESS != (ret = ompi_rte_finalize())) {
|
|
return ret;
|
|
}
|
|
|
|
if (OPAL_SUCCESS != (ret = opal_finalize_util())) {
|
|
return ret;
|
|
}
|
|
|
|
/* All done */
|
|
|
|
return MPI_SUCCESS;
|
|
}
|