1
1
openmpi/ompi/mca/pml/dr/pml_dr_sendreq.h
George Bosilca e361bcb64c Send optimizations.
1. The send path get shorter. The BTL is allowed to return > 0 to specify that the
   descriptor was pushed to the networks, and that the memory attached to it is 
   available again for the upper layer. The MCA_BTL_DES_SEND_ALWAYS_CALLBACK flag
   can be used by the PML to force the BTL to always trigger the callback.
   Unmodified BTL will continue to work as expected, as they will return OMPI_SUCCESS
   which force the PML to have exactly the same behavior as before. Some BTLs have
   been modified: self, sm, tcp, mx.
2. Add send immediate interface to BTL.
   The idea is to have a mechanism of allowing the BTL to take advantage of
   send optimizations such as the ability to deliver data "inline". Some
   network APIs such as Portals allow data to be sent using a "thin" event
   without packing data into a memory descriptor. This interface change
   allows the BTL to use such capabilities and allows for other optimizations
   in the future. All existing BTLs except for Portals and sm have this interface
   set to NULL.

This commit was SVN r18551.
2008-05-30 03:58:39 +00:00

494 строки
29 KiB
C

/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2007 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) 2004-2006 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef OMPI_PML_DR_SEND_REQUEST_H
#define OMPI_PML_DR_SEND_REQUEST_H
#include "opal/util/crc.h"
#include "ompi_config.h"
#include "ompi/datatype/convertor.h"
#include "ompi/mca/btl/btl.h"
#include "ompi/mca/pml/base/pml_base_sendreq.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/bml/bml.h"
#include "ompi/mca/btl/btl.h"
#include "pml_dr_comm.h"
#include "pml_dr_hdr.h"
#include "pml_dr_vfrag.h"
#include "pml_dr_endpoint.h"
#include "opal/event/event.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
struct mca_pml_dr_send_request_t {
mca_pml_base_send_request_t req_send;
mca_pml_dr_comm_proc_t* req_proc;
mca_pml_dr_endpoint_t* req_endpoint;
#if OMPI_HAVE_THREAD_SUPPORT
volatile int32_t req_state;
volatile int32_t req_lock;
#else
int32_t req_state;
int32_t req_lock;
#endif
size_t req_pipeline_depth;
size_t req_bytes_delivered;
size_t req_send_offset;
mca_pml_dr_vfrag_t* req_vfrag;
mca_pml_dr_vfrag_t req_vfrag0;
opal_list_t req_retrans;
mca_btl_base_descriptor_t* req_descriptor; /* descriptor for first frag, retransmission */
};
typedef struct mca_pml_dr_send_request_t mca_pml_dr_send_request_t;
OBJ_CLASS_DECLARATION(mca_pml_dr_send_request_t);
#define MCA_PML_DR_SEND_REQUEST_ALLOC( \
comm, \
dst, \
sendreq, \
rc) \
{ \
ompi_proc_t *proc = ompi_comm_peer_lookup( comm, dst ); \
ompi_free_list_item_t* item; \
\
if(NULL == proc) { \
rc = OMPI_ERR_OUT_OF_RESOURCE; \
} else { \
rc = OMPI_SUCCESS; \
OMPI_FREE_LIST_WAIT(&mca_pml_base_send_requests, item, rc); \
sendreq = (mca_pml_dr_send_request_t*)item; \
sendreq->req_send.req_base.req_proc = proc; \
opal_list_append(&mca_pml_dr.send_active, \
(opal_list_item_t*) sendreq); \
} \
}
#define MCA_PML_DR_SEND_REQUEST_INIT( \
sendreq, \
addr, \
count, \
datatype, \
peer, \
tag, \
comm, \
sendmode, \
persistent) \
do { \
mca_bml_base_endpoint_t* endpoint = \
sendreq->req_send.req_base.req_proc->proc_bml; \
bool do_csum = mca_pml_dr.enable_csum && \
(endpoint->btl_flags_or & MCA_BTL_FLAGS_NEED_CSUM); \
/* increment reference counts */ \
OBJ_RETAIN(comm); \
OBJ_RETAIN(datatype); \
\
OMPI_REQUEST_INIT(&(sendreq)->req_send.req_base.req_ompi, persistent); \
(sendreq)->req_send.req_addr = addr; \
(sendreq)->req_send.req_send_mode = sendmode; \
(sendreq)->req_send.req_base.req_addr = addr; \
(sendreq)->req_send.req_base.req_count = count; \
(sendreq)->req_send.req_base.req_datatype = datatype; \
(sendreq)->req_send.req_base.req_peer = (int32_t)peer; \
(sendreq)->req_send.req_base.req_tag = (int32_t)tag; \
(sendreq)->req_send.req_base.req_comm = comm; \
(sendreq)->req_send.req_base.req_pml_complete = OPAL_INT_TO_BOOL(persistent); \
(sendreq)->req_send.req_base.req_free_called = false; \
(sendreq)->req_send.req_base.req_ompi.req_status._cancelled = 0; \
\
/* initialize datatype convertor for this request */ \
/* if(count > 0) { */ \
/* We will create a convertor specialized for the */ \
/* remote architecture and prepared with the datatype. */ \
ompi_convertor_copy_and_prepare_for_send( \
(sendreq)->req_send.req_base.req_proc->proc_convertor, \
(sendreq)->req_send.req_base.req_datatype, \
(sendreq)->req_send.req_base.req_count, \
(sendreq)->req_send.req_base.req_addr, \
(do_csum ? CONVERTOR_WITH_CHECKSUM: 0), \
&(sendreq)->req_send.req_base.req_convertor ); \
ompi_convertor_get_packed_size(&(sendreq)->req_send.req_base.req_convertor, \
&((sendreq)->req_send.req_bytes_packed) ); \
/* } else { */ \
/* (sendreq)->req_send.req_bytes_packed = 0; */ \
/* } */ \
} while(0)
#define MCA_PML_DR_SEND_REQUEST_START(sendreq, rc) \
do { \
mca_pml_dr_comm_t* comm = sendreq->req_send.req_base.req_comm->c_pml_comm; \
mca_pml_dr_endpoint_t* pml_endpoint = \
(mca_pml_dr_endpoint_t*)sendreq->req_send.req_base.req_proc->proc_pml; \
mca_bml_base_endpoint_t* bml_endpoint = \
sendreq->req_send.req_base.req_proc->proc_bml; \
mca_pml_dr_comm_proc_t* proc = \
comm->procs + sendreq->req_send.req_base.req_peer; \
mca_bml_base_btl_t* bml_btl; \
size_t size = sendreq->req_send.req_bytes_packed; \
size_t eager_limit; \
if(pml_endpoint == NULL || bml_endpoint == NULL) { \
rc = OMPI_ERR_UNREACH; \
break; \
} \
\
bml_btl = mca_bml_base_btl_array_get_next(&bml_endpoint->btl_eager); \
MCA_PML_DR_VFRAG_INIT(&sendreq->req_vfrag0); \
sendreq->req_vfrag0.vf_id = OPAL_THREAD_ADD32(&pml_endpoint->vfrag_seq,1); \
sendreq->req_vfrag0.bml_btl = bml_btl; \
sendreq->req_vfrag = &sendreq->req_vfrag0; \
sendreq->req_endpoint = pml_endpoint; \
assert(pml_endpoint->bml_endpoint == bml_endpoint); \
sendreq->req_proc = proc; \
\
sendreq->req_lock = 0; \
sendreq->req_pipeline_depth = 1; \
sendreq->req_bytes_delivered = 0; \
sendreq->req_state = 0; \
sendreq->req_send_offset = 0; \
sendreq->req_send.req_base.req_pml_complete = false; \
sendreq->req_send.req_base.req_ompi.req_complete = false; \
sendreq->req_send.req_base.req_ompi.req_state = OMPI_REQUEST_ACTIVE; \
sendreq->req_send.req_base.req_ompi.req_status._cancelled = 0; \
sendreq->req_send.req_base.req_sequence = OPAL_THREAD_ADD32(&proc->send_sequence,1); \
\
/* select a btl */ \
eager_limit = bml_btl->btl_eager_limit - sizeof(mca_pml_dr_hdr_t); \
if(size <= eager_limit) { \
switch(sendreq->req_send.req_send_mode) { \
case MCA_PML_BASE_SEND_SYNCHRONOUS: \
rc = mca_pml_dr_send_request_start_rndv(sendreq, bml_btl, size, 0); \
break; \
case MCA_PML_BASE_SEND_BUFFERED: \
rc = mca_pml_dr_send_request_start_copy(sendreq, bml_btl, size); \
break; \
case MCA_PML_BASE_SEND_COMPLETE: \
rc = mca_pml_dr_send_request_start_prepare(sendreq, bml_btl, size); \
break; \
default: \
if (bml_btl->btl_flags & MCA_BTL_FLAGS_SEND_INPLACE) { \
rc = mca_pml_dr_send_request_start_prepare(sendreq, bml_btl, size); \
} else { \
rc = mca_pml_dr_send_request_start_copy(sendreq, bml_btl, size); \
} \
break; \
} \
} else { \
size = eager_limit; \
if(sendreq->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED) { \
rc = mca_pml_dr_send_request_start_buffered(sendreq, bml_btl, size); \
} else { \
rc = mca_pml_dr_send_request_start_rndv(sendreq, bml_btl, size, 0); \
} \
} \
} while (0)
/*
* Mark a send request as completed at the MPI level.
*/
#define MCA_PML_DR_SEND_REQUEST_MPI_COMPLETE(sendreq) \
do { \
(sendreq)->req_send.req_base.req_ompi.req_status.MPI_SOURCE = \
(sendreq)->req_send.req_base.req_comm->c_my_rank; \
(sendreq)->req_send.req_base.req_ompi.req_status.MPI_TAG = \
(sendreq)->req_send.req_base.req_tag; \
(sendreq)->req_send.req_base.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS; \
(sendreq)->req_send.req_base.req_ompi.req_status._count = \
(sendreq)->req_send.req_bytes_packed; \
ompi_request_complete( &((sendreq)->req_send.req_base.req_ompi), true ); \
} while(0)
/*
* The request fini is responsible for releasing all ressources at the PML
* level. It will never be called directly from the upper level, as it should
* only be an internal call to the PML. However, in the case when the user
* already lost the MPI reference to the request (MPI_Request_free was called)
* fini should completely free the MPI request.
*/
#define MCA_PML_DR_SEND_REQUEST_PML_COMPLETE(sendreq) \
do { \
assert( false == sendreq->req_send.req_base.req_pml_complete ); \
opal_list_remove_item(&mca_pml_dr.send_active, \
(opal_list_item_t*) sendreq); \
if (sendreq->req_send.req_send_mode == MCA_PML_BASE_SEND_BUFFERED && \
sendreq->req_send.req_addr != sendreq->req_send.req_base.req_addr) { \
mca_pml_base_bsend_request_fini((ompi_request_t*)sendreq); \
} \
\
if( false == sendreq->req_send.req_base.req_ompi.req_complete ) { \
/* Should only be called for long messages (maybe synchronous) */ \
MCA_PML_DR_SEND_REQUEST_MPI_COMPLETE(sendreq); \
} \
sendreq->req_send.req_base.req_pml_complete = true; \
\
if( sendreq->req_send.req_base.req_free_called ) { \
MCA_PML_DR_SEND_REQUEST_RETURN( sendreq ); \
} else { \
if(sendreq->req_send.req_base.req_ompi.req_persistent && \
(0 != sendreq->req_send.req_bytes_packed) ) { \
/* rewind convertor */ \
size_t offset = 0; \
ompi_convertor_set_position(&sendreq->req_send.req_base.req_convertor, \
&offset); \
} \
} \
} while (0)
/*
* Release resources associated with a request
*/
#define MCA_PML_DR_SEND_REQUEST_RETURN(sendreq) \
do { \
/* Let the base handle the reference counts */ \
MCA_PML_BASE_SEND_REQUEST_FINI((&(sendreq)->req_send)); \
OMPI_FREE_LIST_RETURN( &mca_pml_base_send_requests, \
(ompi_free_list_item_t*)sendreq ); \
} while(0)
/*
* Lookup/allocate a vfrag for the pending send
*/
#define MCA_PML_DR_SEND_REQUEST_VFRAG_INIT(sendreq, pml_endpoint, size, vfrag) \
do { \
size_t max_send_size = pml_endpoint->bml_endpoint->btl_max_send_size - \
sizeof(mca_pml_dr_frag_hdr_t); \
size_t div = size / max_send_size; \
\
MCA_PML_DR_VFRAG_INIT(vfrag); \
if(div == 0) { \
vfrag->vf_len = 1; \
vfrag->vf_size = size; \
vfrag->vf_mask = 1; \
} else if(div > 64) { \
vfrag->vf_len = 64; \
vfrag->vf_size = (max_send_size << 6); /* size * 64 */ \
vfrag->vf_mask = ~(uint64_t)0; \
} else if (div == 64) { \
size_t mod = size % max_send_size; \
vfrag->vf_len = 64; \
vfrag->vf_size = (mod ? (size - mod) : size); \
vfrag->vf_mask = ~(uint64_t)0; \
} else { \
size_t mod = size % max_send_size; \
vfrag->vf_len = div + (mod ? 1 : 0); \
vfrag->vf_size = size; \
if(vfrag->vf_len == 64) \
vfrag->vf_mask = ~(uint64_t)0; \
else \
vfrag->vf_mask = (((uint64_t)1 << vfrag->vf_len) - (uint64_t)1); \
} \
vfrag->vf_id = OPAL_THREAD_ADD32(&pml_endpoint->vfrag_seq,1); \
vfrag->vf_offset = sendreq->req_send_offset; \
vfrag->vf_max_send_size = max_send_size; \
vfrag->vf_send.pval = sendreq; \
sendreq->req_vfrag = vfrag; \
} while(0)
/*
* Reschedule unacked fragments
*/
#define MCA_PML_DR_SEND_REQUEST_VFRAG_RETRANS(sendreq, vfrag) \
do { \
if(((vfrag)->vf_state & MCA_PML_DR_VFRAG_RETRANS) == 0) { \
opal_list_append(&(sendreq)->req_retrans, (opal_list_item_t*)(vfrag));\
(vfrag)->vf_state |= MCA_PML_DR_VFRAG_RETRANS; \
} \
(vfrag)->vf_state &= ~MCA_PML_DR_VFRAG_NACKED; \
(vfrag)->vf_idx = 0; \
} while(0)
/*
* Update bytes delivered on request based on supplied descriptor
*/
#define MCA_PML_DR_SEND_REQUEST_SET_BYTES_DELIVERED(sendreq, vfrag, hdrlen) \
do { \
sendreq->req_bytes_delivered += vfrag->vf_size; \
} while(0)
/*
* Attempt to process any pending requests
*/
#define MCA_PML_DR_SEND_REQUEST_PROCESS_PENDING() \
do { \
/* advance pending requests */ \
while(opal_list_get_size(&mca_pml_dr.send_pending)) { \
mca_pml_dr_send_request_t* sendreq; \
OPAL_THREAD_LOCK(&ompi_request_lock); \
sendreq = (mca_pml_dr_send_request_t*) \
opal_list_remove_first(&mca_pml_dr.send_pending); \
OPAL_THREAD_UNLOCK(&ompi_request_lock); \
if(NULL == sendreq) \
break; \
opal_list_append(&mca_pml_dr.send_active, \
(opal_list_item_t*) sendreq); \
mca_pml_dr_send_request_schedule(sendreq); \
} \
} while (0)
/*
* Requeue first fragment of message for retransmission
*/
#define MCA_PML_DR_SEND_REQUEST_EAGER_RETRY(sendreq, vfrag) \
do { \
mca_btl_base_descriptor_t *des_old, *des_new; \
OPAL_THREAD_ADD64(&vfrag->vf_pending,1); \
MCA_PML_DR_DEBUG(0, \
(0, "%s:%d:%s: retransmitting eager\n", \
__FILE__, __LINE__, __func__)); \
assert(sendreq->req_descriptor->des_src != NULL); \
\
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_pipeline_depth,1); \
OPAL_THREAD_ADD64(&(vfrag)->vf_pending,1); \
(vfrag)->vf_state &= ~MCA_PML_DR_VFRAG_NACKED; \
\
des_old = sendreq->req_descriptor; \
mca_bml_base_alloc(vfrag->bml_btl, &des_new, \
MCA_BTL_NO_ORDER, des_old->des_src->seg_len, \
des_old->des_flags); \
sendreq->req_descriptor = des_new; \
memcpy(des_new->des_src->seg_addr.pval, \
des_old->des_src->seg_addr.pval, \
des_old->des_src->seg_len); \
des_new->des_flags = des_old->des_flags; \
des_new->des_cbdata = des_old->des_cbdata; \
des_new->des_cbfunc = des_old->des_cbfunc; \
mca_bml_base_send(vfrag->bml_btl, des_new, MCA_BTL_TAG_PML); \
} while(0)
/*
* Requeue first fragment of message for retransmission
*/
#define MCA_PML_DR_SEND_REQUEST_RNDV_PROBE(sendreq, vfrag) \
do { \
mca_pml_dr_endpoint_t* endpoint = sendreq->req_endpoint; \
mca_bml_base_btl_t* bml_btl = \
mca_bml_base_btl_array_get_next(&endpoint->bml_endpoint->btl_eager); \
mca_btl_base_descriptor_t *des_old, *des_new; \
mca_pml_dr_hdr_t *hdr; \
bool do_csum = mca_pml_dr.enable_csum && \
(bml_btl->btl_flags & MCA_BTL_FLAGS_NEED_CSUM); \
MCA_PML_DR_DEBUG(0,(0, "%s:%d:%s: (re)transmitting rndv probe\n", \
__FILE__, __LINE__, __func__)); \
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_pipeline_depth,1); \
OPAL_THREAD_ADD64(&vfrag->vf_pending,1); \
(vfrag)->vf_state &= ~MCA_PML_DR_VFRAG_NACKED; \
\
assert(sendreq->req_descriptor->des_src != NULL); \
des_old = sendreq->req_descriptor; \
mca_bml_base_alloc(bml_btl, &des_new, \
MCA_BTL_NO_ORDER, \
sizeof(mca_pml_dr_rendezvous_hdr_t), \
des_old->des_flags); \
/* build hdr */ \
hdr = (mca_pml_dr_hdr_t*)des_new->des_src->seg_addr.pval; \
hdr->hdr_common.hdr_flags = 0; \
hdr->hdr_common.hdr_type = MCA_PML_DR_HDR_TYPE_RNDV; \
hdr->hdr_common.hdr_dst = endpoint->dst; \
hdr->hdr_common.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid; \
hdr->hdr_common.hdr_src = endpoint->src; \
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; \
hdr->hdr_match.hdr_seq = sendreq->req_send.req_base.req_sequence; \
hdr->hdr_match.hdr_src_ptr.pval = &sendreq->req_vfrag0; \
hdr->hdr_match.hdr_csum = OPAL_CSUM_ZERO; \
hdr->hdr_common.hdr_vid = sendreq->req_vfrag0.vf_id; \
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed; \
hdr->hdr_common.hdr_csum = (do_csum ? \
opal_csum(hdr, sizeof(mca_pml_dr_rendezvous_hdr_t)): OPAL_CSUM_ZERO); \
des_new->des_flags = des_old->des_flags; \
des_new->des_cbdata = des_old->des_cbdata; \
des_new->des_cbfunc = des_old->des_cbfunc; \
mca_bml_base_send(bml_btl, des_new, MCA_BTL_TAG_PML); \
} while(0)
/**
* Start the specified request
*/
int mca_pml_dr_send_request_start_buffered(
mca_pml_dr_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl,
size_t size);
int mca_pml_dr_send_request_start_copy(
mca_pml_dr_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl,
size_t size);
int mca_pml_dr_send_request_start_prepare(
mca_pml_dr_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl,
size_t size);
int mca_pml_dr_send_request_start_rndv(
mca_pml_dr_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl,
size_t size,
int flags);
/**
* Schedule additional fragments
*/
int mca_pml_dr_send_request_schedule(
mca_pml_dr_send_request_t* sendreq);
int mca_pml_dr_send_request_reschedule(
mca_pml_dr_send_request_t* sendreq,
mca_pml_dr_vfrag_t* vfrag);
/**
* Acknowledgment of vfrags
*/
void mca_pml_dr_send_request_match_ack(
mca_btl_base_module_t* btl,
mca_pml_dr_ack_hdr_t*);
void mca_pml_dr_send_request_rndv_ack(
mca_btl_base_module_t* btl,
mca_pml_dr_ack_hdr_t*);
void mca_pml_dr_send_request_frag_ack(
mca_btl_base_module_t* btl,
mca_pml_dr_ack_hdr_t*);
void mca_pml_dr_sendreq_cleanup_active(mca_btl_base_module_t* btl);
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif