1
1
openmpi/ompi/mca/btl/mx/btl_mx_component.c
George Bosilca 75a35ed7ee Implement the PUT protocol over MX. The send/receive approach give the best
performance on a 2G Myrinet card, as it look like pipelining the messages
by 1M is faster than a simple send/receive. However, when using a 10G card
the send/receive will limit the maximum bandwidth to 2.5Gbs. The reason is
the scarce bus resources that have to be shared between the Myrinet hardware
and the memcpy operation. The PUT protocol remove the memcpy, we now have a 
true zero-copy mechanism. But, there is no pipelining yet as it look like the
RDMA pipeline somehow disappeared from the OB1 PML ...

This commit was SVN r12925.
2006-12-24 22:52:46 +00:00

468 строки
19 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.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "ompi/constants.h"
#include "opal/event/event.h"
#include "opal/util/opal_environ.h"
#include "opal/util/if.h"
#include "opal/util/argv.h"
#include "opal/util/output.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/btl/btl.h"
#include "opal/mca/base/mca_base_param.h"
#include "ompi/mca/pml/base/pml_base_module_exchange.h"
#include "orte/mca/errmgr/errmgr.h"
#include "ompi/mca/mpool/base/base.h"
#include "btl_mx.h"
#include "btl_mx_frag.h"
#include "btl_mx_endpoint.h"
#include "ompi/mca/btl/base/base.h"
#include "ompi/mca/btl/base/btl_base_error.h"
extern char** environ;
mca_btl_mx_component_t mca_btl_mx_component = {
{
/* First, the mca_base_component_t struct containing meta information
about the component itself */
{
/* Indicate that we are a pml v1.0.0 component (which also implies a
specific MCA version) */
MCA_BTL_BASE_VERSION_1_0_1,
"mx", /* MCA component name */
OMPI_MAJOR_VERSION, /* MCA component major version */
OMPI_MINOR_VERSION, /* MCA component minor version */
OMPI_RELEASE_VERSION, /* MCA component release version */
mca_btl_mx_component_open, /* component open */
mca_btl_mx_component_close /* component close */
},
/* Next the MCA v1.0.0 component meta data */
{
/* Whether the component is checkpointable or not */
false
},
mca_btl_mx_component_init,
mca_btl_mx_component_progress,
}
};
/*
* Called by MCA framework to open the component, registers
* component parameters.
*/
int mca_btl_mx_component_open(void)
{
int tmp;
/* initialize state */
mca_btl_mx_component.mx_num_btls = 0;
mca_btl_mx_component.mx_btls = NULL;
/* initialize objects */
OBJ_CONSTRUCT(&mca_btl_mx_component.mx_procs, opal_list_t);
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "max_btls",
"Maximum number of accepted Myrinet cards",
false, false, 1, &mca_btl_mx_component.mx_max_btls );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "timeout",
"Timeout for connections",
false, false, MX_INFINITE, &mca_btl_mx_component.mx_timeout );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "retries",
"Number of retries for each new connection before considering the peer as unreacheable",
false, false, 20, &mca_btl_mx_component.mx_connection_retries );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "filter",
"Unique ID for the application (used to connect to the peers)",
false, false, 0xdeadbeef, &mca_btl_mx_component.mx_filter );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "self",
"Enable the MX support for self communications",
false, false, 0, &mca_btl_mx_component.mx_support_self );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "shared_mem",
"Enable the MX support for shared memory",
false, false, 0, &mca_btl_mx_component.mx_support_sharedmem );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "free_list_num",
"Number of allocated default request",
false, false, 8, &mca_btl_mx_component.mx_free_list_num );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "free_list_inc",
"Number of request we allocate each time we miss some",
false, false, 32, &mca_btl_mx_component.mx_free_list_inc );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "free_list_max",
"Maximum number of request this device is allowed to allocate",
false, false, 128, &mca_btl_mx_component.mx_free_list_max );
/* The ompi_free_list has a problem if the (max - num) is not
* divisible by the increament. So make sure it is ...
*/
if( (mca_btl_mx_component.mx_free_list_max - mca_btl_mx_component.mx_free_list_num) %
mca_btl_mx_component.mx_free_list_inc ) {
int overhead = (mca_btl_mx_component.mx_free_list_max - mca_btl_mx_component.mx_free_list_num) %
mca_btl_mx_component.mx_free_list_inc;
mca_btl_mx_component.mx_free_list_max -= overhead;
}
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "max_posted_recv",
"Number of received posted in advance. Increasing this number for communication bound application can lead to visible improvement in performances",
false, false, 16, &mca_btl_mx_component.mx_max_posted_recv );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "exclusivity",
"Priority compared with the others devices (used only when several devices are available",
false, false, 50, (int*) &mca_btl_mx_module.super.btl_exclusivity );
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "first_frag_size",
"Size of the first fragment for the rendez-vous protocol over MX",
true, true, 16*1024 - 20, &tmp);
mca_btl_mx_module.super.btl_eager_limit = tmp;
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "min_send_size",
"Minimum send fragment size ...",
false, false, 32*1024 - 40, &tmp);
mca_btl_mx_module.super.btl_min_send_size = tmp;
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "max_send_size",
"Maximum send fragment size withour RDMA ...",
false, false, 128*1024, &tmp);
mca_btl_mx_module.super.btl_max_send_size = tmp;
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "min_rdma_size",
"Minimum size of fragment for the RDMA protocol",
false, false, 1024*1024, &tmp);
mca_btl_mx_module.super.btl_min_rdma_size = tmp;
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "max_rdma_size",
"Maximum size of fragment for the RDMA protocol",
false, false, 1024*1024, &tmp);
mca_btl_mx_module.super.btl_max_rdma_size = tmp;
mca_base_param_reg_int( (mca_base_component_t*)&mca_btl_mx_component, "flags",
"Flags to activate/deactivate the RDMA",
false, false, MCA_BTL_FLAGS_SEND_INPLACE | MCA_BTL_FLAGS_PUT,
(int*)&mca_btl_mx_module.super.btl_flags );
return OMPI_SUCCESS;
}
/*
* component cleanup - sanity checking of queue lengths
*/
int mca_btl_mx_component_close(void)
{
if( NULL == mca_btl_mx_component.mx_btls )
return OMPI_SUCCESS;
mx_finalize();
/* release resources */
OBJ_DESTRUCT(&mca_btl_mx_component.mx_send_eager_frags);
OBJ_DESTRUCT(&mca_btl_mx_component.mx_send_user_frags);
OBJ_DESTRUCT(&mca_btl_mx_component.mx_recv_frags);
OBJ_DESTRUCT(&mca_btl_mx_component.mx_procs);
OBJ_DESTRUCT(&mca_btl_mx_component.mx_pending_acks);
OBJ_DESTRUCT(&mca_btl_mx_component.mx_lock);
return OMPI_SUCCESS;
}
/*
* Create and intialize an MX PTL module, where each module
* represents a specific NIC.
*/
static mca_btl_mx_module_t* mca_btl_mx_create(uint64_t addr)
{
mca_btl_mx_module_t* mx_btl;
mx_return_t status;
uint32_t nic_id;
status = mx_nic_id_to_board_number( addr, &nic_id );
if( MX_SUCCESS != status ) {
return NULL;
}
mx_btl = malloc(sizeof(mca_btl_mx_module_t));
if( NULL == mx_btl ) return NULL;
/* copy over default settings */
memcpy( mx_btl, &mca_btl_mx_module, sizeof(mca_btl_mx_module_t) );
OBJ_CONSTRUCT( &mx_btl->mx_peers, opal_list_t );
OBJ_CONSTRUCT( &mx_btl->mx_lock, opal_mutex_t );
/* open local endpoint */
status = mx_open_endpoint( nic_id, MX_ANY_ENDPOINT,
mca_btl_mx_component.mx_filter,
NULL, 0, &mx_btl->mx_endpoint);
if(status != MX_SUCCESS) {
opal_output(0, "mca_btl_mx_init: mx_open_endpoint() failed with status=%d\n", status);
mca_btl_mx_finalize( &mx_btl->super );
return NULL;
}
/* query the endpoint address */
if((status = mx_get_endpoint_addr( mx_btl->mx_endpoint,
&mx_btl->mx_endpoint_addr)) != MX_SUCCESS) {
opal_output(0, "mca_btl_mx_init: mx_get_endpoint_addr() failed with status=%d\n", status);
mca_btl_mx_finalize( &mx_btl->super );
return NULL;
}
return mx_btl;
}
/*
* MX component initialization:
* - check if MX can be initialized.
* - and construct all static objects.
*/
mca_btl_base_module_t** mca_btl_mx_component_init(int *num_btl_modules,
bool enable_progress_threads,
bool enable_mpi_threads)
{
mca_btl_base_module_t** btls;
mx_return_t status;
uint32_t size, count;
int32_t i;
uint64_t *nic_addrs;
mca_btl_mx_addr_t *mx_addrs;
*num_btl_modules = 0;
if (enable_progress_threads) {
opal_output( 0, "mca_btl_mx_component_init: progress threads requested but not supported");
mca_pml_base_modex_send(&mca_btl_mx_component.super.btl_version,
NULL, 0);
return NULL;
}
/* set the MX error handle to always return. This function is the only MX function
* allowed to be called before mx_init in order to make sure that if the MX is not
* up and running the MX library does not exit the application.
*/
mx_set_error_handler(MX_ERRORS_RETURN);
if( 0 == mca_btl_mx_component.mx_support_sharedmem )
opal_setenv( "MX_DISABLE_SHMEM", "1", true, &environ );
if( 0 == mca_btl_mx_component.mx_support_self )
opal_setenv( "MX_DISABLE_SELF", "1", true, &environ );
/* First check if MX is available ... */
if( MX_SUCCESS != (status = mx_init()) ) {
opal_output( 0, "mca_btl_mx_component_init: mx_init() failed with status = %d (%s)\n",
status, mx_strerror(status) );
mca_pml_base_modex_send(&mca_btl_mx_component.super.btl_version,
NULL, 0);
return NULL;
}
/* initialize objects */
OBJ_CONSTRUCT(&mca_btl_mx_component.mx_send_eager_frags, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_btl_mx_component.mx_send_user_frags, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_btl_mx_component.mx_recv_frags, ompi_free_list_t);
OBJ_CONSTRUCT(&mca_btl_mx_component.mx_procs, opal_list_t);
OBJ_CONSTRUCT(&mca_btl_mx_component.mx_pending_acks, opal_list_t);
OBJ_CONSTRUCT(&mca_btl_mx_component.mx_lock, opal_mutex_t);
ompi_free_list_init( &mca_btl_mx_component.mx_send_eager_frags,
sizeof(mca_btl_mx_frag_t) + mca_btl_mx_module.super.btl_eager_limit,
OBJ_CLASS(mca_btl_mx_frag_t),
mca_btl_mx_component.mx_free_list_num,
mca_btl_mx_component.mx_free_list_max,
mca_btl_mx_component.mx_free_list_inc,
NULL ); /* use default allocator */
ompi_free_list_init( &mca_btl_mx_component.mx_send_user_frags,
sizeof(mca_btl_mx_frag_t),
OBJ_CLASS(mca_btl_mx_frag_t),
mca_btl_mx_component.mx_free_list_num,
mca_btl_mx_component.mx_free_list_max,
mca_btl_mx_component.mx_free_list_inc,
NULL ); /* use default allocator */
ompi_free_list_init( &mca_btl_mx_component.mx_recv_frags,
sizeof(mca_btl_mx_frag_t),
OBJ_CLASS(mca_btl_mx_frag_t),
mca_btl_mx_component.mx_free_list_num,
mca_btl_mx_component.mx_free_list_max,
mca_btl_mx_component.mx_free_list_inc,
NULL ); /* use default allocator */
/* intialize process hash table */
OBJ_CONSTRUCT( &mca_btl_mx_component.mx_procs, opal_list_t );
/* get the number of card available on the system */
if( (status = mx_get_info( NULL, MX_NIC_COUNT, NULL, 0,
&mca_btl_mx_component.mx_num_btls, sizeof(uint32_t))) != MX_SUCCESS ) {
opal_output(0, "mca_btl_mx_component_init: mx_get_info(MX_NIC_COUNT) failed with status=%d\n", status);
mca_pml_base_modex_send(&mca_btl_mx_component.super.btl_version,
NULL, 0);
return NULL;
}
if (0 == mca_btl_mx_component.mx_num_btls) {
mca_btl_base_error_no_nics("Myrinet/MX", "NIC");
mca_pml_base_modex_send(&mca_btl_mx_component.super.btl_version,
NULL, 0);
return NULL;
}
/* check for limit on number of btls */
if(mca_btl_mx_component.mx_num_btls > mca_btl_mx_component.mx_max_btls)
mca_btl_mx_component.mx_num_btls = mca_btl_mx_component.mx_max_btls;
/* Now we know how many NIC are available on the system. We will create a BTL for each one
* and then give a pointer to the BTL to the upper level.
*/
mca_btl_mx_component.mx_btls = malloc( mca_btl_mx_component.mx_num_btls * sizeof(mca_btl_base_module_t*) );
if( NULL == mca_btl_mx_component.mx_btls )
return NULL;
/* determine the NIC ids */
size = sizeof(uint64_t) * (mca_btl_mx_component.mx_num_btls + 1);
if( NULL == (nic_addrs = (uint64_t*)malloc(size)) )
return NULL;
if( (status = mx_get_info( NULL, MX_NIC_IDS, NULL, 0,
nic_addrs, size)) != MX_SUCCESS) {
opal_output(0, "MX BTL error (mx_get_info failed) size = %ld [%s] #cards %d\n",
size, mx_strerror(status), mca_btl_mx_component.mx_num_btls );
free(nic_addrs);
return NULL;
}
size = sizeof(mca_btl_mx_addr_t) * mca_btl_mx_component.mx_num_btls;
mx_addrs = (mca_btl_mx_addr_t*)malloc( size );
if( NULL == mx_addrs ) {
free( nic_addrs );
return NULL;
}
/* create a btl for each NIC */
for( i = count = 0; i < mca_btl_mx_component.mx_num_btls; i++ ) {
mca_btl_mx_module_t* btl = mca_btl_mx_create(nic_addrs[i]);
if( NULL == btl ) {
continue;
}
status = mx_decompose_endpoint_addr( btl->mx_endpoint_addr, &(mx_addrs[i].nic_id),
&(mx_addrs[i].endpoint_id) );
if( MX_SUCCESS != status ) {
OBJ_RELEASE( btl );
continue;
}
mca_btl_mx_component.mx_btls[count++] = btl;
}
size = sizeof(mca_btl_mx_addr_t) * count;
if( 0 == count ) {
/* No active BTL module */
}
mca_btl_mx_component.mx_num_btls = count;
/* publish the MX addresses via the MCA framework */
mca_pml_base_modex_send( &mca_btl_mx_component.super.btl_version, mx_addrs, size );
free( nic_addrs );
free( mx_addrs );
btls = malloc( mca_btl_mx_component.mx_num_btls * sizeof(mca_btl_base_module_t*) );
if( NULL == btls ) {
free( mca_btl_mx_component.mx_btls );
mca_btl_mx_component.mx_num_btls = 0; /* no active BTL modules */
return NULL;
}
memcpy( btls, mca_btl_mx_component.mx_btls,
mca_btl_mx_component.mx_num_btls*sizeof(mca_btl_mx_module_t*) );
*num_btl_modules = mca_btl_mx_component.mx_num_btls;
return btls;
}
/*
* MX component progress.
*/
int mca_btl_mx_component_progress()
{
int32_t num_progressed = 0, i;
mx_status_t mx_status;
mx_return_t mx_return;
mx_segment_t mx_segment;
mx_request_t mx_request;
mca_btl_mx_frag_t* frag;
for( i = 0; i < mca_btl_mx_component.mx_num_btls; i++ ) {
mca_btl_mx_module_t* mx_btl = mca_btl_mx_component.mx_btls[i];
uint32_t mx_result = 0;
/* pre-post receive */
#if 0
if( mx_btl->mx_recvs_posted == 0 ) {
OPAL_THREAD_ADD32( &mx_btl->mx_recvs_posted, 1 );
MCA_BTL_MX_POST( mx_btl, frag );
}
#endif
/*if( mx_btl->mx_posted_request ) { */
mx_return = mx_ipeek( mx_btl->mx_endpoint, &mx_request, &mx_result );
if( mx_return != MX_SUCCESS ) {
opal_output(0, "mca_btl_mx_component_progress: mx_ipeek() failed with status %d\n",
mx_return);
continue;
}
if( mx_result == 0 ) {
continue;
}
mx_return = mx_test( mx_btl->mx_endpoint, &mx_request, &mx_status, &mx_result);
if( mx_return != MX_SUCCESS ) {
opal_output(0, "mca_btl_mx_progress: mx_test() failed with status=%dn",
mx_return);
continue;
}
frag = mx_status.context;
if( 0 == frag->base.des_dst_cnt ) { /* it's a send */
/* call the completion callback */
frag->base.des_cbfunc( &(mx_btl->super), frag->endpoint, &(frag->base), OMPI_SUCCESS);
} else { /* and this one is a receive */
mca_btl_base_recv_reg_t* reg;
reg = &(mx_btl->mx_reg[frag->tag]);
frag->base.des_dst->seg_len = mx_status.msg_length;
reg->cbfunc( &(mx_btl->super), frag->tag, &(frag->base), reg->cbdata );
/*
* The upper level extract the data from the fragment. Now we can register the fragment
* again with the MX BTL.
*/
mx_segment.segment_ptr = frag->base.des_dst->seg_addr.pval;
mx_segment.segment_length = mca_btl_mx_module.super.btl_eager_limit;
mx_return = mx_irecv( mx_btl->mx_endpoint, &mx_segment, 1, (uint64_t)frag->tag,
BTL_MX_RECV_MASK,
frag, &(frag->mx_request) );
if( MX_SUCCESS != mx_return ) {
opal_output( 0, "Fail to re-register a fragment with the MX NIC ...\n" );
}
}
/*MCA_BTL_MX_PROGRESS(mx_btl, mx_status);*/
/*
* on the mx_status we have now the pointer attached to the request. This pointer indicate
* which fragment we are working on. On the status we have the status of the operation, so
* we know what we are supposed to do next.
*/
num_progressed++;
}
return num_progressed;
}