ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

replaced the old CID allocation algorithm with the blocked algorithm. The

Просмотр исходного кода 
impace in the communicator directory is still not great since the interface
for allocating a Cid has not changed..

This commit was SVN r12836.
Этот коммит содержится в:
Mohamad Chaarawi 2006-12-12 22:01:39 +00:00
родитель 10af8ab454
Коммит cae083dec6
3 изменённых файлов: 157 добавлений и 70 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

206
ompi/communicator/comm_cid.c
Просмотреть файл

@ -9,6 +9,7 @@
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006 University of Houston. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@ -30,6 +31,7 @@
#include "ompi/mca/coll/base/base.h"
#include "orte/mca/rml/rml.h"
#include "ompi/request/request.h"
#include "ompi/runtime/mpiruntime.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
@ -41,6 +43,8 @@ extern "C" {
* and a bridge-comm (intercomm-create scenario).
*/
static int cid_block_start = 28;
typedef int ompi_comm_cid_allredfct (int *inbuf, int* outbuf,
int count, struct ompi_op_t *op,
ompi_communicator_t *comm,
@ -104,7 +108,6 @@ static opal_mutex_t ompi_cid_lock;
#endif /* OMPI_HAVE_THREAD_SUPPORT */
static opal_list_t ompi_registered_comms;
int ompi_comm_nextcid ( ompi_communicator_t* newcomm,
ompi_communicator_t* comm,
ompi_communicator_t* bridgecomm,
@ -112,14 +115,9 @@ int ompi_comm_nextcid ( ompi_communicator_t* newcomm,
void* remote_leader,
int mode, int send_first )
{
int nextlocal_cid;
int nextcid;
int done=0;
int response=0, glresponse=0;
int nextcid, block;
int global_block_start;
bool flag;
int start=ompi_mpi_communicators.lowest_free;
int i;
ompi_comm_cid_allredfct* allredfnct;
@ -146,79 +144,149 @@ int ompi_comm_nextcid ( ompi_communicator_t* newcomm,
break;
}
OPAL_THREAD_LOCK(&ompi_cid_lock);
ompi_comm_register_cid (comm->c_contextid);
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
while (!done) {
/**
* This is the real algorithm described in the doc
*/
/**
* In case multi-threading is enabled, we revert to the old algorithm
* starting from cid_block_start
*/
if (MPI_THREAD_MULTIPLE == ompi_mpi_thread_provided) {
int nextlocal_cid;
int done=0;
int response=0, glresponse=0;
int start=ompi_mpi_communicators.lowest_free;
int i;
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;
}
ompi_comm_register_cid (comm->c_contextid);
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
for (i=start; i < mca_pml.pml_max_contextid ; i++) {
flag=ompi_pointer_array_test_and_set_item(&ompi_mpi_communicators, i, comm);
if (true == flag) {
nextlocal_cid = i;
break;
}
}
(allredfnct)(&nextlocal_cid, &nextcid, 1, MPI_MAX, comm, bridgecomm,
local_leader, remote_leader, send_first );
if (nextcid == nextlocal_cid) {
response = 1; /* fine with me */
}
else {
ompi_pointer_array_set_item(&ompi_mpi_communicators,
nextlocal_cid, NULL);
flag = ompi_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 */
}
}
(allredfnct)(&response, &glresponse, 1, MPI_MIN, comm, bridgecomm,
local_leader, remote_leader, send_first );
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 */
ompi_pointer_array_set_item(&ompi_mpi_communicators,
nextcid, NULL);
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=ompi_pointer_array_test_and_set_item(&ompi_mpi_communicators,
i, comm);
if (true == flag) {
nextlocal_cid = i;
break;
}
}
(allredfnct)(&nextlocal_cid, &nextcid, 1, MPI_MAX, comm, bridgecomm,
local_leader, remote_leader, send_first );
if (nextcid == nextlocal_cid) {
response = 1; /* fine with me */
}
else {
ompi_pointer_array_set_item(&ompi_mpi_communicators,
nextlocal_cid, NULL);
flag = ompi_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 */
}
}
(allredfnct)(&response, &glresponse, 1, MPI_MIN, comm, bridgecomm,
local_leader, remote_leader, send_first );
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 */
ompi_pointer_array_set_item(&ompi_mpi_communicators,
nextcid, NULL);
}
start = nextcid+1; /* that's where we can start the next round */
}
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;
ompi_pointer_array_set_item (&ompi_mpi_communicators, nextcid, newcomm);
OPAL_THREAD_LOCK(&ompi_cid_lock);
ompi_comm_unregister_cid (comm->c_contextid);
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
return (MPI_SUCCESS);
}
/**
* In case the communication mode is INTRA_OOB or INTAR_BRIDGE, we use the
* highest-free algorithm
*/
if ( OMPI_COMM_CID_INTRA_OOB == mode || OMPI_COMM_CID_INTRA_BRIDGE == mode) {
(allredfnct)(&cid_block_start, &global_block_start, 1,
MPI_MAX, comm, bridgecomm,
local_leader, remote_leader, send_first );
cid_block_start = global_block_start;
nextcid = cid_block_start;
cid_block_start = cid_block_start + 1;
}
else {
flag=false;
block = 0;
if( 0 == comm->c_contextid ) {
block = OMPI_COMM_BLOCK_WORLD;
}
else {
block = OMPI_COMM_BLOCK_OTHERS;
}
while(!flag) {
/**
* If the communicator has IDs available then allocate one for the child
*/
if(MPI_UNDEFINED != comm->c_id_available &&
MPI_UNDEFINED != comm->c_id_start_index &&
block > comm->c_id_available - comm->c_id_start_index) {
nextcid = comm->c_id_available;
flag=ompi_pointer_array_test_and_set_item (&ompi_mpi_communicators,
nextcid, comm);
}
/**
* Otherwise the communicator needs to negotiate a new block of IDs
*/
else {
(allredfnct)(&cid_block_start, &global_block_start, 1,
MPI_MAX, comm, bridgecomm,
local_leader, remote_leader, send_first );
cid_block_start = global_block_start;
comm->c_id_available = cid_block_start;
comm->c_id_start_index = cid_block_start;
cid_block_start = cid_block_start + block;
}
}
comm->c_id_available++;
}
/* set the according values to the newcomm */
newcomm->c_contextid = nextcid;
newcomm->c_f_to_c_index = newcomm->c_contextid;
ompi_pointer_array_set_item (&ompi_mpi_communicators, nextcid, newcomm);
OPAL_THREAD_LOCK(&ompi_cid_lock);
ompi_comm_unregister_cid (comm->c_contextid);
OPAL_THREAD_UNLOCK(&ompi_cid_lock);
return (MPI_SUCCESS);
}
/**************************************************************************/
/**************************************************************************/
/**************************************************************************/

7
ompi/communicator/comm_init.c
Просмотреть файл

@ -9,6 +9,7 @@
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006 University of Houston. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@ -77,6 +78,8 @@ int ompi_comm_init(void)
ompi_group_increment_proc_count (group);
ompi_mpi_comm_world.c_contextid = 0;
ompi_mpi_comm_world.c_id_start_index = 4;
ompi_mpi_comm_world.c_id_available = 4;
ompi_mpi_comm_world.c_f_to_c_index = 0;
ompi_mpi_comm_world.c_my_rank = group->grp_my_rank;
ompi_mpi_comm_world.c_local_group = group;
@ -108,6 +111,8 @@ int ompi_comm_init(void)
ompi_mpi_comm_self.c_contextid = 1;
ompi_mpi_comm_self.c_f_to_c_index = 1;
ompi_mpi_comm_self.c_id_start_index = 20;
ompi_mpi_comm_self.c_id_available = 20;
ompi_mpi_comm_self.c_my_rank = group->grp_my_rank;
ompi_mpi_comm_self.c_local_group = group;
ompi_mpi_comm_self.c_remote_group = group;
@ -257,6 +262,8 @@ static void ompi_comm_construct(ompi_communicator_t* comm)
comm->c_f_to_c_index = MPI_UNDEFINED;
comm->c_name[0] = '\0';
comm->c_contextid = MPI_UNDEFINED;
comm->c_id_available = MPI_UNDEFINED;
comm->c_id_start_index = MPI_UNDEFINED;
comm->c_flags = 0;
comm->c_my_rank = 0;
comm->c_cube_dim = 0;

14
ompi/communicator/communicator.h
Просмотреть файл

@ -10,6 +10,7 @@
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2006 University of Houston. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@ -83,6 +84,13 @@ OMPI_DECLSPEC OBJ_CLASS_DECLARATION(ompi_communicator_t);
#define OMPI_COMM_CID_INTRA_BRIDGE 0x00000080
#define OMPI_COMM_CID_INTRA_OOB 0x00000100
/**
* The block of CIDs allocated for MPI_COMM_WORLD
* and other communicators
*/
#define OMPI_COMM_BLOCK_WORLD 16
#define OMPI_COMM_BLOCK_OTHERS 8
OMPI_DECLSPEC extern ompi_pointer_array_t ompi_mpi_communicators;
struct ompi_communicator_t {
@ -95,6 +103,11 @@ struct ompi_communicator_t {
uint32_t c_flags; /* flags, e.g. intercomm,
topology, etc. */
int c_id_available; /* the currently available Cid for allocation
to a child*/
int c_id_start_index; /* the starting index of the block of cids
allocated to tthis communicator*/
ompi_group_t *c_local_group;
ompi_group_t *c_remote_group;
@ -344,7 +357,6 @@ struct ompi_communicator_t {
int mode,
int send_first);
/**
* shut down the communicator infrastructure.
*/