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.
912 строки
32 KiB
C
912 строки
32 KiB
C
/* -*- Mode: C; c-basic-offset:4 ; -*- */
|
|
/*
|
|
* Copyright (c) 2004-2005 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) 2004-2008 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) 2007 Cisco Systems, Inc. All rights reserved.
|
|
* Copyright (c) 2007 Voltaire All rights reserved.
|
|
* Copyright (c) 2006-2010 University of Houston. All rights reserved.
|
|
* Copyright (c) 2009 Sun Microsystems, Inc. All rights reserved.
|
|
* Copyright (c) 2012-2013 Los Alamos National Security, LLC. All rights
|
|
* reserved.
|
|
* Copyright (c) 2012 Oak Ridge National Labs. All rights reserved.
|
|
* Copyright (c) 2013 Intel, Inc. All rights reserved.
|
|
* $COPYRIGHT$
|
|
*
|
|
* Additional copyrights may follow
|
|
*
|
|
* $HEADER$
|
|
*/
|
|
|
|
#include "ompi_config.h"
|
|
|
|
#include "opal/dss/dss.h"
|
|
#include "ompi/proc/proc.h"
|
|
#include "ompi/communicator/communicator.h"
|
|
#include "ompi/op/op.h"
|
|
#include "ompi/constants.h"
|
|
#include "opal/class/opal_pointer_array.h"
|
|
#include "opal/class/opal_list.h"
|
|
#include "ompi/mca/pml/pml.h"
|
|
#include "ompi/mca/rte/rte.h"
|
|
#include "ompi/mca/coll/base/base.h"
|
|
#include "ompi/request/request.h"
|
|
#include "ompi/runtime/ompi_module_exchange.h"
|
|
#include "ompi/runtime/mpiruntime.h"
|
|
|
|
BEGIN_C_DECLS
|
|
|
|
/**
|
|
* These functions make sure, that we determine the global result over
|
|
* an intra communicators (simple), an inter-communicator and a
|
|
* pseudo inter-communicator described by two separate intra-comms
|
|
* and a bridge-comm (intercomm-create scenario).
|
|
*/
|
|
|
|
|
|
typedef int ompi_comm_cid_allredfct (int *inbuf, int* outbuf,
|
|
int count, struct ompi_op_t *op,
|
|
ompi_communicator_t *comm,
|
|
ompi_communicator_t *bridgecomm,
|
|
void* lleader, void* rleader,
|
|
int send_first );
|
|
|
|
static int ompi_comm_allreduce_intra (int *inbuf, int* outbuf,
|
|
int count, struct ompi_op_t *op,
|
|
ompi_communicator_t *intercomm,
|
|
ompi_communicator_t *bridgecomm,
|
|
void* local_leader,
|
|
void* remote_ledaer,
|
|
int send_first );
|
|
|
|
static int ompi_comm_allreduce_inter (int *inbuf, int *outbuf,
|
|
int count, struct ompi_op_t *op,
|
|
ompi_communicator_t *intercomm,
|
|
ompi_communicator_t *bridgecomm,
|
|
void* local_leader,
|
|
void* remote_leader,
|
|
int send_first );
|
|
|
|
static int ompi_comm_allreduce_intra_bridge(int *inbuf, int* outbuf,
|
|
int count, struct ompi_op_t *op,
|
|
ompi_communicator_t *intercomm,
|
|
ompi_communicator_t *bridgecomm,
|
|
void* local_leader,
|
|
void* remote_leader,
|
|
int send_first);
|
|
|
|
static int ompi_comm_allreduce_intra_oob (int *inbuf, int* outbuf,
|
|
int count, struct ompi_op_t *op,
|
|
ompi_communicator_t *intercomm,
|
|
ompi_communicator_t *bridgecomm,
|
|
void* local_leader,
|
|
void* remote_leader,
|
|
int send_first );
|
|
|
|
static int ompi_comm_register_cid (uint32_t contextid);
|
|
static int ompi_comm_unregister_cid (uint32_t contextid);
|
|
static uint32_t ompi_comm_lowest_cid ( void );
|
|
|
|
struct ompi_comm_reg_t{
|
|
opal_list_item_t super;
|
|
uint32_t cid;
|
|
};
|
|
typedef struct ompi_comm_reg_t ompi_comm_reg_t;
|
|
OMPI_DECLSPEC OBJ_CLASS_DECLARATION(ompi_comm_reg_t);
|
|
|
|
static void ompi_comm_reg_constructor(ompi_comm_reg_t *regcom);
|
|
static void ompi_comm_reg_destructor(ompi_comm_reg_t *regcom);
|
|
|
|
OBJ_CLASS_INSTANCE (ompi_comm_reg_t,
|
|
opal_list_item_t,
|
|
ompi_comm_reg_constructor,
|
|
ompi_comm_reg_destructor );
|
|
|
|
static opal_mutex_t ompi_cid_lock;
|
|
static opal_list_t ompi_registered_comms;
|
|
|
|
|
|
/* This variable is zero (false) if all processes in MPI_COMM_WORLD
|
|
* did not require MPI_THREAD_MULTIPLE support, and is 1 (true) as
|
|
* soon as at least one process requested support for THREAD_MULTIPLE */
|
|
static int ompi_comm_world_thread_level_mult=0;
|
|
|
|
|
|
int ompi_comm_cid_init (void)
|
|
{
|
|
ompi_proc_t **procs, *thisproc;
|
|
uint8_t thread_level;
|
|
void *tlpointer;
|
|
int ret;
|
|
size_t i, size, numprocs;
|
|
|
|
/** Note that the following call only returns processes
|
|
* with the same jobid. This is on purpose, since
|
|
* we switch for the dynamic communicators anyway
|
|
* to the original (slower) cid allocation algorithm.
|
|
*/
|
|
procs = ompi_proc_world ( &numprocs );
|
|
|
|
for ( i=0; i<numprocs; i++ ) {
|
|
thisproc = procs[i];
|
|
|
|
ret = ompi_modex_recv_string("MPI_THREAD_LEVEL", thisproc, &tlpointer, &size);
|
|
if (OMPI_SUCCESS == ret) {
|
|
thread_level = *((uint8_t *) tlpointer);
|
|
if ( OMPI_THREADLEVEL_IS_MULTIPLE (thread_level) ) {
|
|
ompi_comm_world_thread_level_mult = 1;
|
|
break;
|
|
}
|
|
} else if (OMPI_ERR_NOT_IMPLEMENTED == ret) {
|
|
if (ompi_mpi_thread_multiple) {
|
|
ompi_comm_world_thread_level_mult = 1;
|
|
}
|
|
break;
|
|
} else {
|
|
return ret;
|
|
}
|
|
}
|
|
free(procs);
|
|
|
|
return OMPI_SUCCESS;
|
|
}
|
|
|
|
|
|
int ompi_comm_nextcid ( ompi_communicator_t* newcomm,
|
|
ompi_communicator_t* comm,
|
|
ompi_communicator_t* bridgecomm,
|
|
void* local_leader,
|
|
void* remote_leader,
|
|
int mode, int send_first )
|
|
{
|
|
int ret;
|
|
int nextcid;
|
|
bool flag;
|
|
int nextlocal_cid;
|
|
int done=0;
|
|
int response, glresponse=0;
|
|
int start;
|
|
unsigned int i;
|
|
|
|
ompi_comm_cid_allredfct* allredfnct;
|
|
|
|
/**
|
|
* Determine which implementation of allreduce we have to use
|
|
* for the current scenario
|
|
*/
|
|
|
|
switch (mode)
|
|
{
|
|
case OMPI_COMM_CID_INTRA:
|
|
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra;
|
|
break;
|
|
case OMPI_COMM_CID_INTER:
|
|
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_inter;
|
|
break;
|
|
case OMPI_COMM_CID_INTRA_BRIDGE:
|
|
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_bridge;
|
|
break;
|
|
case OMPI_COMM_CID_INTRA_OOB:
|
|
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_oob;
|
|
break;
|
|
default:
|
|
return MPI_UNDEFINED;
|
|
break;
|
|
}
|
|
|
|
do {
|
|
/* Only one communicator function allowed in same time on the
|
|
* same communicator.
|
|
*/
|
|
OPAL_THREAD_LOCK(&ompi_cid_lock);
|
|
response = ompi_comm_register_cid (comm->c_contextid);
|
|
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
|
|
} while (OMPI_SUCCESS != response );
|
|
start = ompi_mpi_communicators.lowest_free;
|
|
|
|
while (!done) {
|
|
/**
|
|
* This is the real algorithm described in the doc
|
|
*/
|
|
OPAL_THREAD_LOCK(&ompi_cid_lock);
|
|
if (comm->c_contextid != ompi_comm_lowest_cid() ) {
|
|
/* if not lowest cid, we do not continue, but sleep and try again */
|
|
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
|
|
continue;
|
|
}
|
|
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
|
|
|
|
for (i=start; i < mca_pml.pml_max_contextid ; i++) {
|
|
flag = opal_pointer_array_test_and_set_item(&ompi_mpi_communicators,
|
|
i, comm);
|
|
if (true == flag) {
|
|
nextlocal_cid = i;
|
|
break;
|
|
}
|
|
}
|
|
|
|
ret = (allredfnct)(&nextlocal_cid, &nextcid, 1, MPI_MAX, comm, bridgecomm,
|
|
local_leader, remote_leader, send_first );
|
|
if( OMPI_SUCCESS != ret ) {
|
|
opal_pointer_array_set_item(&ompi_mpi_communicators, nextlocal_cid, NULL);
|
|
goto release_and_return;
|
|
}
|
|
if (nextcid == nextlocal_cid) {
|
|
response = 1; /* fine with me */
|
|
}
|
|
else {
|
|
opal_pointer_array_set_item(&ompi_mpi_communicators,
|
|
nextlocal_cid, NULL);
|
|
|
|
flag = opal_pointer_array_test_and_set_item(&ompi_mpi_communicators,
|
|
nextcid, comm );
|
|
if (true == flag) {
|
|
response = 1; /* works as well */
|
|
}
|
|
else {
|
|
response = 0; /* nope, not acceptable */
|
|
}
|
|
}
|
|
|
|
ret = (allredfnct)(&response, &glresponse, 1, MPI_MIN, comm, bridgecomm,
|
|
local_leader, remote_leader, send_first );
|
|
if( OMPI_SUCCESS != ret ) {
|
|
opal_pointer_array_set_item(&ompi_mpi_communicators, nextcid, NULL);
|
|
goto release_and_return;
|
|
}
|
|
if (1 == glresponse) {
|
|
done = 1; /* we are done */
|
|
break;
|
|
}
|
|
else if ( 0 == glresponse ) {
|
|
if ( 1 == response ) {
|
|
/* we could use that, but other don't agree */
|
|
opal_pointer_array_set_item(&ompi_mpi_communicators,
|
|
nextcid, NULL);
|
|
}
|
|
start = nextcid+1; /* that's where we can start the next round */
|
|
}
|
|
}
|
|
|
|
/* set the according values to the newcomm */
|
|
newcomm->c_contextid = nextcid;
|
|
newcomm->c_f_to_c_index = newcomm->c_contextid;
|
|
opal_pointer_array_set_item (&ompi_mpi_communicators, nextcid, newcomm);
|
|
|
|
release_and_return:
|
|
OPAL_THREAD_LOCK(&ompi_cid_lock);
|
|
ompi_comm_unregister_cid (comm->c_contextid);
|
|
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
|
|
|
|
return ret;
|
|
}
|
|
|
|
/**************************************************************************/
|
|
/**************************************************************************/
|
|
/**************************************************************************/
|
|
static void ompi_comm_reg_constructor (ompi_comm_reg_t *regcom)
|
|
{
|
|
regcom->cid=MPI_UNDEFINED;
|
|
}
|
|
|
|
static void ompi_comm_reg_destructor (ompi_comm_reg_t *regcom)
|
|
{
|
|
}
|
|
|
|
void ompi_comm_reg_init (void)
|
|
{
|
|
OBJ_CONSTRUCT(&ompi_registered_comms, opal_list_t);
|
|
OBJ_CONSTRUCT(&ompi_cid_lock, opal_mutex_t);
|
|
}
|
|
|
|
void ompi_comm_reg_finalize (void)
|
|
{
|
|
OBJ_DESTRUCT(&ompi_registered_comms);
|
|
OBJ_DESTRUCT(&ompi_cid_lock);
|
|
}
|
|
|
|
|
|
static int ompi_comm_register_cid (uint32_t cid )
|
|
{
|
|
opal_list_item_t *item;
|
|
ompi_comm_reg_t *regcom;
|
|
ompi_comm_reg_t *newentry = OBJ_NEW(ompi_comm_reg_t);
|
|
|
|
newentry->cid = cid;
|
|
if ( !(opal_list_is_empty (&ompi_registered_comms)) ) {
|
|
for (item = opal_list_get_first(&ompi_registered_comms);
|
|
item != opal_list_get_end(&ompi_registered_comms);
|
|
item = opal_list_get_next(item)) {
|
|
regcom = (ompi_comm_reg_t *)item;
|
|
if ( regcom->cid > cid ) {
|
|
break;
|
|
}
|
|
#if OMPI_ENABLE_THREAD_MULTIPLE
|
|
if( regcom->cid == cid ) {
|
|
/**
|
|
* The MPI standard state that is the user responsability to
|
|
* schedule the global communications in order to avoid any
|
|
* kind of troubles. As, managing communicators involve several
|
|
* collective communications, we should enforce a sequential
|
|
* execution order. This test only allow one communicator
|
|
* creation function based on the same communicator.
|
|
*/
|
|
OBJ_RELEASE(newentry);
|
|
return OMPI_ERROR;
|
|
}
|
|
#endif /* OMPI_ENABLE_THREAD_MULTIPLE */
|
|
}
|
|
opal_list_insert_pos (&ompi_registered_comms, item,
|
|
(opal_list_item_t *)newentry);
|
|
}
|
|
else {
|
|
opal_list_append (&ompi_registered_comms, (opal_list_item_t *)newentry);
|
|
}
|
|
|
|
return OMPI_SUCCESS;
|
|
}
|
|
|
|
static int ompi_comm_unregister_cid (uint32_t cid)
|
|
{
|
|
ompi_comm_reg_t *regcom;
|
|
opal_list_item_t *item;
|
|
|
|
for (item = opal_list_get_first(&ompi_registered_comms);
|
|
item != opal_list_get_end(&ompi_registered_comms);
|
|
item = opal_list_get_next(item)) {
|
|
regcom = (ompi_comm_reg_t *)item;
|
|
if(regcom->cid == cid) {
|
|
opal_list_remove_item(&ompi_registered_comms, item);
|
|
OBJ_RELEASE(regcom);
|
|
break;
|
|
}
|
|
}
|
|
return OMPI_SUCCESS;
|
|
}
|
|
|
|
static uint32_t ompi_comm_lowest_cid (void)
|
|
{
|
|
ompi_comm_reg_t *regcom=NULL;
|
|
opal_list_item_t *item=opal_list_get_first (&ompi_registered_comms);
|
|
|
|
regcom = (ompi_comm_reg_t *)item;
|
|
return regcom->cid;
|
|
}
|
|
/**************************************************************************/
|
|
/**************************************************************************/
|
|
/**************************************************************************/
|
|
/* This routine serves two purposes:
|
|
* - the allreduce acts as a kind of Barrier,
|
|
* which avoids, that we have incoming fragments
|
|
* on the new communicator before everybody has set
|
|
* up the comm structure.
|
|
* - some components (e.g. the collective MagPIe component
|
|
* might want to generate new communicators and communicate
|
|
* using the new comm. Thus, it can just be called after
|
|
* the 'barrier'.
|
|
*
|
|
* The reason that this routine is in comm_cid and not in
|
|
* comm.c is, that this file contains the allreduce implementations
|
|
* which are required, and thus we avoid having duplicate code...
|
|
*/
|
|
int ompi_comm_activate ( ompi_communicator_t** newcomm,
|
|
ompi_communicator_t* comm,
|
|
ompi_communicator_t* bridgecomm,
|
|
void* local_leader,
|
|
void* remote_leader,
|
|
int mode,
|
|
int send_first )
|
|
{
|
|
int ret = 0;
|
|
|
|
int ok=0, gok=0;
|
|
ompi_comm_cid_allredfct* allredfnct;
|
|
|
|
/* Step 1: the barrier, after which it is allowed to
|
|
* send messages over the new communicator
|
|
*/
|
|
switch (mode)
|
|
{
|
|
case OMPI_COMM_CID_INTRA:
|
|
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra;
|
|
break;
|
|
case OMPI_COMM_CID_INTER:
|
|
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_inter;
|
|
break;
|
|
case OMPI_COMM_CID_INTRA_BRIDGE:
|
|
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_bridge;
|
|
break;
|
|
case OMPI_COMM_CID_INTRA_OOB:
|
|
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_oob;
|
|
break;
|
|
default:
|
|
return MPI_UNDEFINED;
|
|
break;
|
|
}
|
|
|
|
if (MPI_UNDEFINED != (*newcomm)->c_local_group->grp_my_rank) {
|
|
|
|
/* Initialize the PML stuff in the newcomm */
|
|
if ( OMPI_SUCCESS != (ret = MCA_PML_CALL(add_comm(*newcomm))) ) {
|
|
goto bail_on_error;
|
|
}
|
|
OMPI_COMM_SET_PML_ADDED(*newcomm);
|
|
}
|
|
|
|
|
|
ret = (allredfnct)(&ok, &gok, 1, MPI_MIN, comm, bridgecomm,
|
|
local_leader, remote_leader, send_first );
|
|
if( OMPI_SUCCESS != ret ) {
|
|
goto bail_on_error;
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
* Check to see if this process is in the new communicator.
|
|
*
|
|
* Specifically, this function is invoked by all proceses in the
|
|
* old communicator, regardless of whether they are in the new
|
|
* communicator or not. This is because it is far simpler to use
|
|
* MPI collective functions on the old communicator to determine
|
|
* some data for the new communicator (e.g., remote_leader) than
|
|
* to kludge up our own pseudo-collective routines over just the
|
|
* processes in the new communicator. Hence, *all* processes in
|
|
* the old communicator need to invoke this function.
|
|
*
|
|
* That being said, only processes in the new communicator need to
|
|
* select a coll module for the new communicator. More
|
|
* specifically, proceses who are not in the new communicator
|
|
* should *not* select a coll module -- for example,
|
|
* ompi_comm_rank(newcomm) returns MPI_UNDEFINED for processes who
|
|
* are not in the new communicator. This can cause errors in the
|
|
* selection / initialization of a coll module. Plus, it's
|
|
* wasteful -- processes in the new communicator will end up
|
|
* freeing the new communicator anyway, so we might as well leave
|
|
* the coll selection as NULL (the coll base comm unselect code
|
|
* handles that case properly).
|
|
*/
|
|
if (MPI_UNDEFINED == (*newcomm)->c_local_group->grp_my_rank) {
|
|
return OMPI_SUCCESS;
|
|
}
|
|
|
|
/* Let the collectives components fight over who will do
|
|
collective on this new comm. */
|
|
if (OMPI_SUCCESS != (ret = mca_coll_base_comm_select(*newcomm))) {
|
|
goto bail_on_error;
|
|
}
|
|
|
|
/* For an inter communicator, we have to deal with the potential
|
|
* problem of what is happening if the local_comm that we created
|
|
* has a lower CID than the parent comm. This is not a problem
|
|
* as long as the user calls MPI_Comm_free on the inter communicator.
|
|
* However, if the communicators are not freed by the user but released
|
|
* by Open MPI in MPI_Finalize, we walk through the list of still available
|
|
* communicators and free them one by one. Thus, local_comm is freed before
|
|
* the actual inter-communicator. However, the local_comm pointer in the
|
|
* inter communicator will still contain the 'previous' address of the local_comm
|
|
* and thus this will lead to a segmentation violation. In order to prevent
|
|
* that from happening, we increase the reference counter local_comm
|
|
* by one if its CID is lower than the parent. We cannot increase however
|
|
* its reference counter if the CID of local_comm is larger than
|
|
* the CID of the inter communicators, since a regular MPI_Comm_free would
|
|
* leave in that the case the local_comm hanging around and thus we would not
|
|
* recycle CID's properly, which was the reason and the cause for this trouble.
|
|
*/
|
|
if ( OMPI_COMM_IS_INTER(*newcomm)) {
|
|
if ( OMPI_COMM_CID_IS_LOWER(*newcomm, comm)) {
|
|
OMPI_COMM_SET_EXTRA_RETAIN (*newcomm);
|
|
OBJ_RETAIN (*newcomm);
|
|
}
|
|
}
|
|
|
|
|
|
return OMPI_SUCCESS;
|
|
|
|
bail_on_error:
|
|
OBJ_RELEASE(*newcomm);
|
|
*newcomm = MPI_COMM_NULL;
|
|
return ret;
|
|
}
|
|
|
|
/**************************************************************************/
|
|
/**************************************************************************/
|
|
/**************************************************************************/
|
|
/* Arguments not used in this implementation:
|
|
* - bridgecomm
|
|
* - local_leader
|
|
* - remote_leader
|
|
* - send_first
|
|
*/
|
|
static int ompi_comm_allreduce_intra ( int *inbuf, int *outbuf,
|
|
int count, struct ompi_op_t *op,
|
|
ompi_communicator_t *comm,
|
|
ompi_communicator_t *bridgecomm,
|
|
void* local_leader,
|
|
void* remote_leader,
|
|
int send_first )
|
|
{
|
|
return comm->c_coll.coll_allreduce ( inbuf, outbuf, count, MPI_INT,
|
|
op,comm,
|
|
comm->c_coll.coll_allreduce_module );
|
|
}
|
|
|
|
/* Arguments not used in this implementation:
|
|
* - bridgecomm
|
|
* - local_leader
|
|
* - remote_leader
|
|
* - send_first
|
|
*/
|
|
static int ompi_comm_allreduce_inter ( int *inbuf, int *outbuf,
|
|
int count, struct ompi_op_t *op,
|
|
ompi_communicator_t *intercomm,
|
|
ompi_communicator_t *bridgecomm,
|
|
void* local_leader,
|
|
void* remote_leader,
|
|
int send_first )
|
|
{
|
|
int local_rank, rsize;
|
|
int i, rc;
|
|
int *sbuf;
|
|
int *tmpbuf=NULL;
|
|
int *rcounts=NULL, scount=0;
|
|
int *rdisps=NULL;
|
|
|
|
if ( &ompi_mpi_op_sum.op != op && &ompi_mpi_op_prod.op != op &&
|
|
&ompi_mpi_op_max.op != op && &ompi_mpi_op_min.op != op ) {
|
|
return MPI_ERR_OP;
|
|
}
|
|
|
|
if ( !OMPI_COMM_IS_INTER (intercomm)) {
|
|
return MPI_ERR_COMM;
|
|
}
|
|
|
|
/* Allocate temporary arrays */
|
|
rsize = ompi_comm_remote_size (intercomm);
|
|
local_rank = ompi_comm_rank ( intercomm );
|
|
|
|
tmpbuf = (int *) malloc ( count * sizeof(int));
|
|
rdisps = (int *) calloc ( rsize, sizeof(int));
|
|
rcounts = (int *) calloc ( rsize, sizeof(int) );
|
|
if ( OPAL_UNLIKELY (NULL == tmpbuf || NULL == rdisps || NULL == rcounts)) {
|
|
rc = OMPI_ERR_OUT_OF_RESOURCE;
|
|
goto exit;
|
|
}
|
|
|
|
/* Execute the inter-allreduce: the result of our group will
|
|
be in the buffer of the remote group */
|
|
rc = intercomm->c_coll.coll_allreduce ( inbuf, tmpbuf, count, MPI_INT,
|
|
op, intercomm,
|
|
intercomm->c_coll.coll_allreduce_module);
|
|
if ( OMPI_SUCCESS != rc ) {
|
|
goto exit;
|
|
}
|
|
|
|
if ( 0 == local_rank ) {
|
|
MPI_Request req;
|
|
|
|
/* for the allgatherv later */
|
|
scount = count;
|
|
|
|
/* local leader exchange their data and determine the overall result
|
|
for both groups */
|
|
rc = MCA_PML_CALL(irecv (outbuf, count, MPI_INT, 0,
|
|
OMPI_COMM_ALLREDUCE_TAG,
|
|
intercomm, &req));
|
|
if ( OMPI_SUCCESS != rc ) {
|
|
goto exit;
|
|
}
|
|
rc = MCA_PML_CALL(send (tmpbuf, count, MPI_INT, 0,
|
|
OMPI_COMM_ALLREDUCE_TAG,
|
|
MCA_PML_BASE_SEND_STANDARD,
|
|
intercomm));
|
|
if ( OMPI_SUCCESS != rc ) {
|
|
goto exit;
|
|
}
|
|
rc = ompi_request_wait ( &req, MPI_STATUS_IGNORE );
|
|
if ( OMPI_SUCCESS != rc ) {
|
|
goto exit;
|
|
}
|
|
|
|
if ( &ompi_mpi_op_max.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
if (tmpbuf[i] > outbuf[i]) {
|
|
outbuf[i] = tmpbuf[i];
|
|
}
|
|
}
|
|
}
|
|
else if ( &ompi_mpi_op_min.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
if (tmpbuf[i] < outbuf[i]) {
|
|
outbuf[i] = tmpbuf[i];
|
|
}
|
|
}
|
|
}
|
|
else if ( &ompi_mpi_op_sum.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
outbuf[i] += tmpbuf[i];
|
|
}
|
|
}
|
|
else if ( &ompi_mpi_op_prod.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
outbuf[i] *= tmpbuf[i];
|
|
}
|
|
}
|
|
}
|
|
|
|
/* distribute the overall result to all processes in the other group.
|
|
Instead of using bcast, we are using here allgatherv, to avoid the
|
|
possible deadlock. Else, we need an algorithm to determine,
|
|
which group sends first in the inter-bcast and which receives
|
|
the result first.
|
|
*/
|
|
rcounts[0] = count;
|
|
sbuf = outbuf;
|
|
rc = intercomm->c_coll.coll_allgatherv (sbuf, scount, MPI_INT, outbuf,
|
|
rcounts, rdisps, MPI_INT,
|
|
intercomm,
|
|
intercomm->c_coll.coll_allgatherv_module);
|
|
|
|
exit:
|
|
if ( NULL != tmpbuf ) {
|
|
free ( tmpbuf );
|
|
}
|
|
if ( NULL != rcounts ) {
|
|
free ( rcounts );
|
|
}
|
|
if ( NULL != rdisps ) {
|
|
free ( rdisps );
|
|
}
|
|
|
|
return (rc);
|
|
}
|
|
|
|
/* Arguments not used in this implementation:
|
|
* - send_first
|
|
*/
|
|
static int ompi_comm_allreduce_intra_bridge (int *inbuf, int *outbuf,
|
|
int count, struct ompi_op_t *op,
|
|
ompi_communicator_t *comm,
|
|
ompi_communicator_t *bcomm,
|
|
void* lleader, void* rleader,
|
|
int send_first )
|
|
{
|
|
int *tmpbuf=NULL;
|
|
int local_rank;
|
|
int i;
|
|
int rc;
|
|
int local_leader, remote_leader;
|
|
|
|
local_leader = (*((int*)lleader));
|
|
remote_leader = (*((int*)rleader));
|
|
|
|
if ( &ompi_mpi_op_sum.op != op && &ompi_mpi_op_prod.op != op &&
|
|
&ompi_mpi_op_max.op != op && &ompi_mpi_op_min.op != op ) {
|
|
return MPI_ERR_OP;
|
|
}
|
|
|
|
local_rank = ompi_comm_rank ( comm );
|
|
tmpbuf = (int *) malloc ( count * sizeof(int));
|
|
if ( NULL == tmpbuf ) {
|
|
rc = OMPI_ERR_OUT_OF_RESOURCE;
|
|
goto exit;
|
|
}
|
|
|
|
/* Intercomm_create */
|
|
rc = comm->c_coll.coll_allreduce ( inbuf, tmpbuf, count, MPI_INT,
|
|
op, comm, comm->c_coll.coll_allreduce_module );
|
|
if ( OMPI_SUCCESS != rc ) {
|
|
goto exit;
|
|
}
|
|
|
|
if (local_rank == local_leader ) {
|
|
MPI_Request req;
|
|
|
|
rc = MCA_PML_CALL(irecv ( outbuf, count, MPI_INT, remote_leader,
|
|
OMPI_COMM_ALLREDUCE_TAG,
|
|
bcomm, &req));
|
|
if ( OMPI_SUCCESS != rc ) {
|
|
goto exit;
|
|
}
|
|
rc = MCA_PML_CALL(send (tmpbuf, count, MPI_INT, remote_leader,
|
|
OMPI_COMM_ALLREDUCE_TAG,
|
|
MCA_PML_BASE_SEND_STANDARD, bcomm));
|
|
if ( OMPI_SUCCESS != rc ) {
|
|
goto exit;
|
|
}
|
|
rc = ompi_request_wait_all ( 1, &req, MPI_STATUS_IGNORE);
|
|
if ( OMPI_SUCCESS != rc ) {
|
|
goto exit;
|
|
}
|
|
|
|
if ( &ompi_mpi_op_max.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
if (tmpbuf[i] > outbuf[i]) {
|
|
outbuf[i] = tmpbuf[i];
|
|
}
|
|
}
|
|
}
|
|
else if ( &ompi_mpi_op_min.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
if (tmpbuf[i] < outbuf[i]) {
|
|
outbuf[i] = tmpbuf[i];
|
|
}
|
|
}
|
|
}
|
|
else if ( &ompi_mpi_op_sum.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
outbuf[i] += tmpbuf[i];
|
|
}
|
|
}
|
|
else if ( &ompi_mpi_op_prod.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
outbuf[i] *= tmpbuf[i];
|
|
}
|
|
}
|
|
}
|
|
|
|
rc = comm->c_coll.coll_bcast ( outbuf, count, MPI_INT, local_leader,
|
|
comm, comm->c_coll.coll_bcast_module );
|
|
|
|
exit:
|
|
if (NULL != tmpbuf ) {
|
|
free (tmpbuf);
|
|
}
|
|
|
|
return (rc);
|
|
}
|
|
|
|
typedef struct {
|
|
opal_buffer_t buf;
|
|
bool active;
|
|
} comm_cid_return_t;
|
|
|
|
static void comm_cid_recv(int status,
|
|
ompi_process_name_t* peer,
|
|
opal_buffer_t* buffer,
|
|
ompi_rml_tag_t tag,
|
|
void* cbdata)
|
|
{
|
|
comm_cid_return_t *rcid = (comm_cid_return_t*)cbdata;
|
|
|
|
opal_dss.copy_payload(&rcid->buf, buffer);
|
|
rcid->active = false;
|
|
}
|
|
|
|
/* Arguments not used in this implementation:
|
|
* - bridgecomm
|
|
*
|
|
* lleader is the local rank of root in comm
|
|
* rleader is the OOB contact information of the
|
|
* root processes in the other world.
|
|
*/
|
|
static int ompi_comm_allreduce_intra_oob (int *inbuf, int *outbuf,
|
|
int count, struct ompi_op_t *op,
|
|
ompi_communicator_t *comm,
|
|
ompi_communicator_t *bridgecomm,
|
|
void* lleader, void* rleader,
|
|
int send_first )
|
|
{
|
|
int *tmpbuf=NULL;
|
|
int i;
|
|
int rc;
|
|
int local_leader, local_rank;
|
|
ompi_process_name_t *remote_leader=NULL;
|
|
int32_t size_count;
|
|
comm_cid_return_t rcid;
|
|
|
|
local_leader = (*((int*)lleader));
|
|
remote_leader = (ompi_process_name_t*)rleader;
|
|
size_count = count;
|
|
|
|
if ( &ompi_mpi_op_sum.op != op && &ompi_mpi_op_prod.op != op &&
|
|
&ompi_mpi_op_max.op != op && &ompi_mpi_op_min.op != op ) {
|
|
return MPI_ERR_OP;
|
|
}
|
|
|
|
|
|
local_rank = ompi_comm_rank ( comm );
|
|
tmpbuf = (int *) malloc ( count * sizeof(int));
|
|
if ( NULL == tmpbuf ) {
|
|
rc = OMPI_ERR_OUT_OF_RESOURCE;
|
|
goto exit;
|
|
}
|
|
|
|
/* comm is an intra-communicator */
|
|
rc = comm->c_coll.coll_allreduce(inbuf,tmpbuf,count,MPI_INT,op, comm,
|
|
comm->c_coll.coll_allreduce_module);
|
|
if ( OMPI_SUCCESS != rc ) {
|
|
goto exit;
|
|
}
|
|
|
|
if (local_rank == local_leader ) {
|
|
opal_buffer_t *sbuf;
|
|
|
|
sbuf = OBJ_NEW(opal_buffer_t);
|
|
|
|
if (OPAL_SUCCESS != (rc = opal_dss.pack(sbuf, tmpbuf, (int32_t)count, OPAL_INT))) {
|
|
goto exit;
|
|
}
|
|
|
|
if ( send_first ) {
|
|
if (0 > (rc = ompi_rte_send_buffer_nb(remote_leader, sbuf,
|
|
OMPI_RML_TAG_COMM_CID_INTRA,
|
|
ompi_rte_send_cbfunc, NULL))) {
|
|
goto exit;
|
|
}
|
|
OBJ_CONSTRUCT(&rcid.buf, opal_buffer_t);
|
|
rcid.active = true;
|
|
ompi_rte_recv_buffer_nb(remote_leader, OMPI_RML_TAG_COMM_CID_INTRA,
|
|
OMPI_RML_NON_PERSISTENT, comm_cid_recv, &rcid);
|
|
while (rcid.active) {
|
|
opal_progress();
|
|
}
|
|
}
|
|
else {
|
|
OBJ_CONSTRUCT(&rcid.buf, opal_buffer_t);
|
|
rcid.active = true;
|
|
ompi_rte_recv_buffer_nb(remote_leader, OMPI_RML_TAG_COMM_CID_INTRA,
|
|
OMPI_RML_NON_PERSISTENT, comm_cid_recv, &rcid);
|
|
while (rcid.active) {
|
|
opal_progress();
|
|
}
|
|
if (0 > (rc = ompi_rte_send_buffer_nb(remote_leader, sbuf,
|
|
OMPI_RML_TAG_COMM_CID_INTRA,
|
|
ompi_rte_send_cbfunc, NULL))) {
|
|
goto exit;
|
|
}
|
|
}
|
|
|
|
if (OPAL_SUCCESS != (rc = opal_dss.unpack(&rcid.buf, outbuf, &size_count, OPAL_INT))) {
|
|
goto exit;
|
|
}
|
|
OBJ_DESTRUCT(&rcid.buf);
|
|
count = (int)size_count;
|
|
|
|
if ( &ompi_mpi_op_max.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
if (tmpbuf[i] > outbuf[i]) {
|
|
outbuf[i] = tmpbuf[i];
|
|
}
|
|
}
|
|
}
|
|
else if ( &ompi_mpi_op_min.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
if (tmpbuf[i] < outbuf[i]) {
|
|
outbuf[i] = tmpbuf[i];
|
|
}
|
|
}
|
|
}
|
|
else if ( &ompi_mpi_op_sum.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
outbuf[i] += tmpbuf[i];
|
|
}
|
|
}
|
|
else if ( &ompi_mpi_op_prod.op == op ) {
|
|
for ( i = 0 ; i < count; i++ ) {
|
|
outbuf[i] *= tmpbuf[i];
|
|
}
|
|
}
|
|
}
|
|
|
|
rc = comm->c_coll.coll_bcast (outbuf, count, MPI_INT,
|
|
local_leader, comm,
|
|
comm->c_coll.coll_bcast_module);
|
|
|
|
exit:
|
|
if (NULL != tmpbuf ) {
|
|
free (tmpbuf);
|
|
}
|
|
|
|
return (rc);
|
|
}
|
|
|
|
END_C_DECLS
|