ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
91a281080a
openmpi/ompi/mca/pml/ob1/pml_ob1_recvreq.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

465 строки
20 KiB
C
Исходник Ответственный История

/*
* 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) 2004-2007 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) 2008 UT-Battelle, LLC. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef OMPI_PML_OB1_RECV_REQUEST_H
#define OMPI_PML_OB1_RECV_REQUEST_H
#include "pml_ob1.h"
#include "pml_ob1_rdma.h"
#include "pml_ob1_rdmafrag.h"
#include "ompi/proc/proc.h"
#include "ompi/mca/pml/ob1/pml_ob1_comm.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/pml/base/pml_base_recvreq.h"
#include "ompi/datatype/datatype.h"
BEGIN_C_DECLS
struct mca_pml_ob1_recv_request_t {
mca_pml_base_recv_request_t req_recv;
ompi_ptr_t req_send;
int32_t req_lock;
size_t req_pipeline_depth;
size_t req_bytes_received;
size_t req_bytes_delivered;
size_t req_rdma_offset;
size_t req_send_offset;
uint32_t req_rdma_cnt;
uint32_t req_rdma_idx;
bool req_pending;
bool req_ack_sent; /**< whether ack was sent to the sender */
bool req_match_received; /**< Prevent request to be completed prematurely */
opal_mutex_t lock;
mca_pml_ob1_com_btl_t req_rdma[1];
};
typedef struct mca_pml_ob1_recv_request_t mca_pml_ob1_recv_request_t;
OBJ_CLASS_DECLARATION(mca_pml_ob1_recv_request_t);
static inline bool lock_recv_request(mca_pml_ob1_recv_request_t *recvreq)
{
return OPAL_THREAD_ADD32(&recvreq->req_lock, 1) == 1;
}
static inline bool unlock_recv_request(mca_pml_ob1_recv_request_t *recvreq)
{
return OPAL_THREAD_ADD32(&recvreq->req_lock, -1) == 0;
}
/**
* Allocate a recv request from the modules free list.
*
* @param rc (OUT) OMPI_SUCCESS or error status on failure.
* @return Receive request.
*/
#define MCA_PML_OB1_RECV_REQUEST_ALLOC(recvreq, rc) \
do { \
ompi_free_list_item_t* item; \
rc = OMPI_SUCCESS; \
OMPI_FREE_LIST_GET(&mca_pml_base_recv_requests, item, rc); \
recvreq = (mca_pml_ob1_recv_request_t*)item; \
} while(0)
/**
* Initialize a receive request with call parameters.
*
* @param request (IN) Receive request.
* @param addr (IN) User buffer.
* @param count (IN) Number of elements of indicated datatype.
* @param datatype (IN) User defined datatype.
* @param src (IN) Source rank w/in the communicator.
* @param tag (IN) User defined tag.
* @param comm (IN) Communicator.
* @param persistent (IN) Is this a ersistent request.
*/
#define MCA_PML_OB1_RECV_REQUEST_INIT( request, \
addr, \
count, \
datatype, \
src, \
tag, \
comm, \
persistent) \
do { \
MCA_PML_BASE_RECV_REQUEST_INIT( &(request)->req_recv, \
addr, \
count, \
datatype, \
src, \
tag, \
comm, \
persistent); \
} while(0)
/**
* Mark the request as completed at MPI level for internal purposes.
*
* @param recvreq (IN) Receive request.
*/
#define MCA_PML_OB1_RECV_REQUEST_MPI_COMPLETE( recvreq ) \
do { \
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_COMPLETE, \
&(recvreq->req_recv.req_base), PERUSE_RECV ); \
ompi_request_complete( &(recvreq->req_recv.req_base.req_ompi), true ); \
} while (0)
/*
* Free the PML receive request
*/
#define MCA_PML_OB1_RECV_REQUEST_RETURN(recvreq) \
{ \
MCA_PML_BASE_RECV_REQUEST_FINI(&(recvreq)->req_recv); \
OMPI_FREE_LIST_RETURN( &mca_pml_base_recv_requests, \
(ompi_free_list_item_t*)(recvreq)); \
}
/**
* Complete receive request. Request structure cannot be accessed after calling
* this function any more.
*
* @param recvreq (IN) Receive request.
*/
void static inline
recv_request_pml_complete(mca_pml_ob1_recv_request_t *recvreq)
{
size_t i;
assert(false == recvreq->req_recv.req_base.req_pml_complete);
if(recvreq->req_recv.req_bytes_packed > 0) {
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_END,
&recvreq->req_recv.req_base, PERUSE_RECV );
}
for(i = 0; i < recvreq->req_rdma_cnt; i++) {
mca_mpool_base_registration_t* btl_reg = recvreq->req_rdma[i].btl_reg;
if( NULL != btl_reg && btl_reg->mpool != NULL) {
btl_reg->mpool->mpool_deregister( btl_reg->mpool, btl_reg );
}
}
recvreq->req_rdma_cnt = 0;
OPAL_THREAD_LOCK(&ompi_request_lock);
if(true == recvreq->req_recv.req_base.req_free_called) {
MCA_PML_OB1_RECV_REQUEST_RETURN(recvreq);
} else {
/* initialize request status */
recvreq->req_recv.req_base.req_pml_complete = true;
recvreq->req_recv.req_base.req_ompi.req_status._count =
(int)recvreq->req_bytes_received;
if (recvreq->req_bytes_received > recvreq->req_bytes_delivered) {
recvreq->req_recv.req_base.req_ompi.req_status._count =
(int)recvreq->req_bytes_delivered;
recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR =
MPI_ERR_TRUNCATE;
}
MCA_PML_OB1_RECV_REQUEST_MPI_COMPLETE(recvreq);
}
OPAL_THREAD_UNLOCK(&ompi_request_lock);
}
bool static inline
recv_request_pml_complete_check(mca_pml_ob1_recv_request_t *recvreq)
{
opal_atomic_rmb();
if(recvreq->req_match_received &&
recvreq->req_bytes_received >= recvreq->req_recv.req_bytes_packed &&
lock_recv_request(recvreq)) {
recv_request_pml_complete(recvreq);
return true;
}
return false;
}
extern void mca_pml_ob1_recv_req_start(mca_pml_ob1_recv_request_t *req);
#define MCA_PML_OB1_RECV_REQUEST_START(r) mca_pml_ob1_recv_req_start(r)
static inline void prepare_recv_req_converter(mca_pml_ob1_recv_request_t *req)
{
ompi_convertor_copy_and_prepare_for_recv(
req->req_recv.req_base.req_proc->proc_convertor,
req->req_recv.req_base.req_datatype,
req->req_recv.req_base.req_count,
req->req_recv.req_base.req_addr,
0,
&req->req_recv.req_base.req_convertor);
ompi_convertor_get_unpacked_size(&req->req_recv.req_base.req_convertor,
&req->req_bytes_delivered);
}
#define MCA_PML_OB1_RECV_REQUEST_MATCHED(request, hdr) \
recv_req_matched(request, hdr)
static inline void recv_req_matched(mca_pml_ob1_recv_request_t *req,
mca_pml_ob1_match_hdr_t *hdr)
{
req->req_recv.req_base.req_ompi.req_status.MPI_SOURCE = hdr->hdr_src;
req->req_recv.req_base.req_ompi.req_status.MPI_TAG = hdr->hdr_tag;
req->req_match_received = true;
opal_atomic_wmb();
if(req->req_recv.req_bytes_packed > 0) {
if(MPI_ANY_SOURCE == req->req_recv.req_base.req_peer) {
/* non wildcard prepared during post recv */
prepare_recv_req_converter(req);
}
PERUSE_TRACE_COMM_EVENT(PERUSE_COMM_REQ_XFER_BEGIN,
&req->req_recv.req_base, PERUSE_RECV);
}
}
/**
*
*/
#define MCA_PML_OB1_RECV_REQUEST_UNPACK( \
request, \
segments, \
num_segments, \
seg_offset, \
data_offset, \
bytes_received, \
bytes_delivered) \
do { \
bytes_delivered = 0; \
if(request->req_recv.req_bytes_packed > 0) { \
struct iovec iov[MCA_BTL_DES_MAX_SEGMENTS]; \
uint32_t iov_count = 0; \
size_t max_data = bytes_received; \
size_t n, offset = seg_offset; \
mca_btl_base_segment_t* segment = segments; \
\
OPAL_THREAD_LOCK(&request->lock); \
for( n = 0; n < num_segments; n++, segment++ ) { \
if(offset >= segment->seg_len) { \
offset -= segment->seg_len; \
} else { \
iov[iov_count].iov_len = segment->seg_len - offset; \
iov[iov_count].iov_base = (IOVBASE_TYPE*)((unsigned char*)segment->seg_addr.pval + offset); \
iov_count++; \
} \
} \
PERUSE_TRACE_COMM_OMPI_EVENT (PERUSE_COMM_REQ_XFER_CONTINUE, \
&(recvreq->req_recv.req_base), max_data, \
PERUSE_RECV); \
ompi_convertor_set_position( &(request->req_recv.req_base.req_convertor), \
&data_offset ); \
ompi_convertor_unpack( &(request)->req_recv.req_base.req_convertor, \
iov, \
&iov_count, \
&max_data ); \
bytes_delivered = max_data; \
OPAL_THREAD_UNLOCK(&request->lock); \
} \
} while (0)
/**
*
*/
#define MCA_PML_OB1_RECV_REQUEST_UNPACK_ALL( \
request, \
segments, \
num_segments, \
seg_offset, \
data_offset, \
bytes_received, \
bytes_delivered) \
do { \
bytes_delivered = 0; \
if(request->req_recv.req_bytes_packed > 0) { \
struct iovec iov[MCA_BTL_DES_MAX_SEGMENTS]; \
uint32_t iov_count = 0; \
size_t max_data = bytes_received; \
size_t n, offset = seg_offset; \
mca_btl_base_segment_t* segment = segments; \
\
OPAL_THREAD_LOCK(&request->lock); \
for( n = 0; n < num_segments; n++, segment++ ) { \
if(offset >= segment->seg_len) { \
offset -= segment->seg_len; \
} else { \
iov[iov_count].iov_len = segment->seg_len - offset; \
iov[iov_count].iov_base = (IOVBASE_TYPE*)((unsigned char*)segment->seg_addr.pval + offset); \
iov_count++; \
} \
} \
PERUSE_TRACE_COMM_OMPI_EVENT (PERUSE_COMM_REQ_XFER_CONTINUE, \
&(recvreq->req_recv.req_base), \
bytes_received, \
PERUSE_RECV); \
ompi_convertor_set_position( &(request->req_recv.req_base.req_convertor), \
&data_offset ); \
ompi_convertor_unpack( &(request)->req_recv.req_base.req_convertor, \
iov, \
&iov_count, \
&max_data ); \
bytes_delivered = max_data; \
OPAL_THREAD_UNLOCK(&request->lock); \
} \
} while (0)
/**
*
*/
void mca_pml_ob1_recv_request_progress_match(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
void mca_pml_ob1_recv_request_progress_frag(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
void mca_pml_ob1_recv_request_progress_rndv(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
void mca_pml_ob1_recv_request_progress_rget(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
void mca_pml_ob1_recv_request_matched_probe(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
int mca_pml_ob1_recv_request_schedule_once(
mca_pml_ob1_recv_request_t* req, mca_bml_base_btl_t* start_bml_btl);
static inline int mca_pml_ob1_recv_request_schedule_exclusive(
mca_pml_ob1_recv_request_t* req,
mca_bml_base_btl_t* start_bml_btl)
{
int rc;
do {
rc = mca_pml_ob1_recv_request_schedule_once(req, start_bml_btl);
if(rc == OMPI_ERR_OUT_OF_RESOURCE)
break;
} while(!unlock_recv_request(req));
if(OMPI_SUCCESS == rc)
recv_request_pml_complete_check(req);
return rc;
}
static inline void mca_pml_ob1_recv_request_schedule(
mca_pml_ob1_recv_request_t* req,
mca_bml_base_btl_t* start_bml_btl)
{
if(!lock_recv_request(req))
return;
(void)mca_pml_ob1_recv_request_schedule_exclusive(req, start_bml_btl);
}
#define MCA_PML_OB1_ADD_ACK_TO_PENDING(P, S, D, O) \
do { \
mca_pml_ob1_pckt_pending_t *_pckt; \
int _rc; \
\
MCA_PML_OB1_PCKT_PENDING_ALLOC(_pckt,_rc); \
_pckt->hdr.hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_ACK; \
_pckt->hdr.hdr_ack.hdr_src_req.lval = (S); \
_pckt->hdr.hdr_ack.hdr_dst_req.pval = (D); \
_pckt->hdr.hdr_ack.hdr_send_offset = (O); \
_pckt->proc = (P); \
_pckt->bml_btl = NULL; \
OPAL_THREAD_LOCK(&mca_pml_ob1.lock); \
opal_list_append(&mca_pml_ob1.pckt_pending, \
(opal_list_item_t*)_pckt); \
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); \
} while(0)
int mca_pml_ob1_recv_request_ack_send_btl(ompi_proc_t* proc,
mca_bml_base_btl_t* bml_btl, uint64_t hdr_src_req, void *hdr_dst_req,
uint64_t hdr_rdma_offset, bool nordma);
static inline int mca_pml_ob1_recv_request_ack_send(ompi_proc_t* proc,
uint64_t hdr_src_req, void *hdr_dst_req, uint64_t hdr_send_offset,
bool nordma)
{
size_t i;
mca_bml_base_btl_t* bml_btl;
mca_bml_base_endpoint_t* endpoint =
(mca_bml_base_endpoint_t*)proc->proc_bml;
for(i = 0; i < mca_bml_base_btl_array_get_size(&endpoint->btl_eager); i++) {
bml_btl = mca_bml_base_btl_array_get_next(&endpoint->btl_eager);
if(mca_pml_ob1_recv_request_ack_send_btl(proc, bml_btl, hdr_src_req,
hdr_dst_req, hdr_send_offset, nordma) == OMPI_SUCCESS)
return OMPI_SUCCESS;
}
MCA_PML_OB1_ADD_ACK_TO_PENDING(proc, hdr_src_req, hdr_dst_req,
hdr_send_offset);
return OMPI_ERR_OUT_OF_RESOURCE;
}
int mca_pml_ob1_recv_request_get_frag(mca_pml_ob1_rdma_frag_t* frag);
/* This function tries to continue recvreq that stuck due to resource
* unavailability. Recvreq is added to recv_pending list if scheduling of put
* operation cannot be accomplished for some reason. */
void mca_pml_ob1_recv_request_process_pending(void);
END_C_DECLS
#endif
Показать git blame Копировать постоянную ссылку