1
1
openmpi/ompi/mca/pml/base/pml_base_bsend.c
Rainer Keller 6c5532072a - Split the datatype engine into two parts: an MPI specific part in
OMPI
   and a language agnostic part in OPAL. The convertor is completely
   moved into OPAL.  This offers several benefits as described in RFC
   http://www.open-mpi.org/community/lists/devel/2009/07/6387.php
   namely:
    - Fewer basic types (int* and float* types, boolean and wchar
    - Fixing naming scheme to ompi-nomenclature.
    - Usability outside of the ompi-layer.
 - Due to the fixed nature of simple opal types, their information is
   completely
   known at compile time and therefore constified
 - With fewer datatypes (22), the actual sizes of bit-field types may be
   reduced
   from 64 to 32 bits, allowing reorganizing the opal_datatype
   structure, eliminating holes and keeping data required in convertor
   (upon send/recv) in one cacheline...
   This has implications to the convertor-datastructure and other parts
   of the code.
 - Several performance tests have been run, the netpipe latency does not
   change with
   this patch on Linux/x86-64 on the smoky cluster.
 - Extensive tests have been done to verify correctness (no new
   regressions) using:
   1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and
    ompi-ddt:
    a. running both trunk and ompi-ddt resulted in no differences
       (except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run
       correctly).
    b. with --enable-memchecker and running under valgrind (one buglet
       when run with static found in test-suite, commited)
   2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt:
      all passed (except for the dynamic/ tests failed!! as trunk/MTT)
   3. compilation and usage of HDF5 tests on Jaguar using PGI and
      PathScale compilers.
   4. compilation and usage on Scicortex.
 - Please note, that for the heterogeneous case, (-m32 compiled
   binaries/ompi), neither
   ompi-trunk, nor ompi-ddt branch would successfully launch.

This commit was SVN r21641.
2009-07-13 04:56:31 +00:00

394 строки
12 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) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2007 Sun Microsystems, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "opal/threads/mutex.h"
#include "opal/threads/condition.h"
#include "ompi/datatype/ompi_datatype.h"
#include "ompi/mca/allocator/base/base.h"
#include "ompi/mca/allocator/allocator.h"
#include "opal/mca/base/mca_base_param.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/pml/base/pml_base_request.h"
#include "ompi/mca/pml/base/pml_base_sendreq.h"
#include "ompi/mca/pml/base/pml_base_bsend.h"
#include "ompi/mca/mpool/mpool.h"
#ifdef HAVE_UNISTD_H
#include <unistd.h>
#endif /* HAVE_UNISTD_H */
static opal_mutex_t mca_pml_bsend_mutex; /* lock for thread safety */
static opal_condition_t mca_pml_bsend_condition; /* condition variable to block on detach */
static mca_allocator_base_component_t* mca_pml_bsend_allocator_component;
static mca_allocator_base_module_t* mca_pml_bsend_allocator; /* sub-allocator to manage users buffer */
static size_t mca_pml_bsend_usersize; /* user provided buffer size */
static unsigned char *mca_pml_bsend_userbase; /* user provided buffer base */
static unsigned char *mca_pml_bsend_base; /* adjusted base of user buffer */
static unsigned char *mca_pml_bsend_addr; /* current offset into user buffer */
static size_t mca_pml_bsend_size; /* adjusted size of user buffer */
static size_t mca_pml_bsend_count; /* number of outstanding requests */
static size_t mca_pml_bsend_pagesz; /* mmap page size */
static int mca_pml_bsend_pagebits; /* number of bits in pagesz */
static int32_t mca_pml_bsend_init = 0;
/*
* Routine to return pages to sub-allocator as needed
*/
static void* mca_pml_bsend_alloc_segment(
struct mca_mpool_base_module_t* module,
size_t* size_inout,
mca_mpool_base_registration_t** registration)
{
void *addr;
size_t size = *size_inout;
if(mca_pml_bsend_addr + size > mca_pml_bsend_base + mca_pml_bsend_size) {
return NULL;
}
/* allocate all that is left */
size = mca_pml_bsend_size - (mca_pml_bsend_addr - mca_pml_bsend_base);
addr = mca_pml_bsend_addr;
mca_pml_bsend_addr += size;
*size_inout = size;
if (NULL != registration) *registration = NULL;
return addr;
}
/*
* One time initialization at startup
*/
int mca_pml_base_bsend_init(bool thread_safe)
{
int id = mca_base_param_register_string("pml", "base", "bsend_allocator", NULL, "basic");
char *name;
size_t tmp;
if(OPAL_THREAD_ADD32(&mca_pml_bsend_init, 1) > 1)
return OMPI_SUCCESS;
/* initialize static objects */
OBJ_CONSTRUCT(&mca_pml_bsend_mutex, opal_mutex_t);
OBJ_CONSTRUCT(&mca_pml_bsend_condition, opal_condition_t);
/* lookup name of the allocator to use for buffered sends */
mca_base_param_lookup_string(id, &name);
if(NULL == (mca_pml_bsend_allocator_component = mca_allocator_component_lookup(name))) {
free(name);
return OMPI_ERR_BUFFER;
}
free(name);
/* determine page size */
tmp = mca_pml_bsend_pagesz = sysconf(_SC_PAGESIZE);
mca_pml_bsend_pagebits = 0;
while( tmp != 0 ) {
tmp >>= 1;
mca_pml_bsend_pagebits++;
}
return OMPI_SUCCESS;
}
/*
* One-time cleanup at shutdown - release any resources.
*/
int mca_pml_base_bsend_fini(void)
{
if(OPAL_THREAD_ADD32(&mca_pml_bsend_init,-1) > 0)
return OMPI_SUCCESS;
if(NULL != mca_pml_bsend_allocator)
mca_pml_bsend_allocator->alc_finalize(mca_pml_bsend_allocator);
mca_pml_bsend_allocator = NULL;
OBJ_DESTRUCT(&mca_pml_bsend_condition);
OBJ_DESTRUCT(&mca_pml_bsend_mutex);
return OMPI_SUCCESS;
}
/*
* User-level call to attach buffer.
*/
int mca_pml_base_bsend_attach(void* addr, int size)
{
int align;
bool thread_safe = ompi_mpi_thread_multiple;
if(NULL == addr || size <= 0) {
return OMPI_ERR_BUFFER;
}
/* check for buffer already attached */
OPAL_THREAD_LOCK(&mca_pml_bsend_mutex);
if(NULL != mca_pml_bsend_allocator) {
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_ERR_BUFFER;
}
/* try to create an instance of the allocator - to determine thread safety level */
mca_pml_bsend_allocator = mca_pml_bsend_allocator_component->allocator_init(thread_safe, mca_pml_bsend_alloc_segment, NULL, NULL);
if(NULL == mca_pml_bsend_allocator) {
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_ERR_BUFFER;
}
/*
* Save away what the user handed in. This is done in case the
* base and size are modified for alignment issues.
*/
mca_pml_bsend_userbase = (unsigned char*)addr;
mca_pml_bsend_usersize = size;
/*
* Align to pointer boundaries. The bsend overhead is large enough
* to account for this. Compute any alignment that needs to be done.
*/
align = sizeof(void *) - ((size_t)addr & (sizeof(void *) - 1));
/* setup local variables */
mca_pml_bsend_base = (unsigned char *)addr + align;
mca_pml_bsend_addr = (unsigned char *)addr + align;
mca_pml_bsend_size = size - align;
mca_pml_bsend_count = 0;
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_SUCCESS;
}
/*
* User-level call to detach buffer
*/
int mca_pml_base_bsend_detach(void* addr, int* size)
{
OPAL_THREAD_LOCK(&mca_pml_bsend_mutex);
/* is buffer attached */
if(NULL == mca_pml_bsend_allocator) {
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_ERR_BUFFER;
}
/* wait on any pending requests */
while(mca_pml_bsend_count != 0)
opal_condition_wait(&mca_pml_bsend_condition, &mca_pml_bsend_mutex);
/* free resources associated with the allocator */
mca_pml_bsend_allocator->alc_finalize(mca_pml_bsend_allocator);
mca_pml_bsend_allocator = NULL;
/* return current settings */
if(NULL != addr)
*((void**)addr) = mca_pml_bsend_userbase;
if(NULL != size)
*size = (int)mca_pml_bsend_usersize;
/* reset local variables */
mca_pml_bsend_userbase = NULL;
mca_pml_bsend_usersize = 0;
mca_pml_bsend_base = NULL;
mca_pml_bsend_addr = NULL;
mca_pml_bsend_size = 0;
mca_pml_bsend_count = 0;
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_SUCCESS;
}
/*
* pack send buffer into buffer
*/
int mca_pml_base_bsend_request_start(ompi_request_t* request)
{
mca_pml_base_send_request_t* sendreq = (mca_pml_base_send_request_t*)request;
struct iovec iov;
unsigned int iov_count;
size_t max_data;
int rc;
if(sendreq->req_bytes_packed > 0) {
/* has a buffer been provided */
OPAL_THREAD_LOCK(&mca_pml_bsend_mutex);
if(NULL == mca_pml_bsend_addr) {
sendreq->req_addr = NULL;
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_ERR_BUFFER;
}
/* allocate a buffer to hold packed message */
sendreq->req_addr = mca_pml_bsend_allocator->alc_alloc(
mca_pml_bsend_allocator, sendreq->req_bytes_packed, 0, NULL);
if(NULL == sendreq->req_addr) {
/* release resources when request is freed */
sendreq->req_base.req_pml_complete = true;
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_ERR_BUFFER;
}
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
/* The convertor is already initialized in the begining so we just have to
* pack the data in the newly allocated buffer.
*/
iov.iov_base = (IOVBASE_TYPE*)sendreq->req_addr;
iov.iov_len = sendreq->req_bytes_packed;
iov_count = 1;
max_data = iov.iov_len;
if((rc = opal_convertor_pack( &sendreq->req_base.req_convertor,
&iov,
&iov_count,
&max_data )) < 0) {
return OMPI_ERROR;
}
/* setup convertor to point to packed buffer (at position zero) */
opal_convertor_prepare_for_send( &sendreq->req_base.req_convertor, &(ompi_mpi_packed.dt.super),
max_data, sendreq->req_addr );
/* increment count of pending requests */
mca_pml_bsend_count++;
}
return OMPI_SUCCESS;
}
/*
* allocate buffer
*/
int mca_pml_base_bsend_request_alloc(ompi_request_t* request)
{
mca_pml_base_send_request_t* sendreq = (mca_pml_base_send_request_t*)request;
assert( sendreq->req_bytes_packed > 0 );
/* has a buffer been provided */
OPAL_THREAD_LOCK(&mca_pml_bsend_mutex);
if(NULL == mca_pml_bsend_addr) {
sendreq->req_addr = NULL;
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_ERR_BUFFER;
}
/* allocate a buffer to hold packed message */
sendreq->req_addr = mca_pml_bsend_allocator->alc_alloc(
mca_pml_bsend_allocator, sendreq->req_bytes_packed, 0, NULL);
if(NULL == sendreq->req_addr) {
/* release resources when request is freed */
sendreq->req_base.req_pml_complete = true;
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
/* progress communications, with the hope that more resources
* will be freed */
opal_progress();
return OMPI_ERR_BUFFER;
}
/* increment count of pending requests */
mca_pml_bsend_count++;
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_SUCCESS;
}
/*
* allocate buffer
*/
void* mca_pml_base_bsend_request_alloc_buf( size_t length )
{
void* buf = NULL;
/* has a buffer been provided */
OPAL_THREAD_LOCK(&mca_pml_bsend_mutex);
if(NULL == mca_pml_bsend_addr) {
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return NULL;
}
/* allocate a buffer to hold packed message */
buf = mca_pml_bsend_allocator->alc_alloc(
mca_pml_bsend_allocator, length, 0, NULL);
if(NULL == buf) {
/* release resources when request is freed */
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
/* progress communications, with the hope that more resources
* will be freed */
opal_progress();
return NULL;
}
/* increment count of pending requests */
mca_pml_bsend_count++;
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return buf;
}
/*
* Request completed - free buffer and decrement pending count
*/
int mca_pml_base_bsend_request_free(void* addr)
{
/* remove from list of pending requests */
OPAL_THREAD_LOCK(&mca_pml_bsend_mutex);
/* free buffer */
mca_pml_bsend_allocator->alc_free(mca_pml_bsend_allocator, addr);
/* decrement count of buffered requests */
if(--mca_pml_bsend_count == 0)
opal_condition_signal(&mca_pml_bsend_condition);
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_SUCCESS;
}
/*
* Request completed - free buffer and decrement pending count
*/
int mca_pml_base_bsend_request_fini(ompi_request_t* request)
{
mca_pml_base_send_request_t* sendreq = (mca_pml_base_send_request_t*)request;
if(sendreq->req_bytes_packed == 0 ||
sendreq->req_addr == NULL ||
sendreq->req_addr == sendreq->req_base.req_addr)
return OMPI_SUCCESS;
/* remove from list of pending requests */
OPAL_THREAD_LOCK(&mca_pml_bsend_mutex);
/* free buffer */
mca_pml_bsend_allocator->alc_free(mca_pml_bsend_allocator, sendreq->req_addr);
sendreq->req_addr = sendreq->req_base.req_addr;
/* decrement count of buffered requests */
if(--mca_pml_bsend_count == 0)
opal_condition_signal(&mca_pml_bsend_condition);
OPAL_THREAD_UNLOCK(&mca_pml_bsend_mutex);
return OMPI_SUCCESS;
}