1
1

handle the situation where you have an upper liomit on the simultanious number of pending aio operations.

Этот коммит содержится в:
Edgar Gabriel 2015-01-29 10:23:36 -06:00
родитель 2d79963cb3
Коммит 5d58ce2113
4 изменённых файлов: 90 добавлений и 13 удалений

Просмотреть файл

@ -9,7 +9,7 @@
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2008-2014 University of Houston. All rights reserved.
* Copyright (c) 2008-2015 University of Houston. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@ -32,6 +32,8 @@
#include <aio.h>
#endif
int fbtl_posix_max_aio_active_reqs=2048;
#include "ompi/mca/fbtl/fbtl.h"
#include "ompi/mca/fbtl/posix/fbtl_posix.h"
@ -95,6 +97,13 @@ int mca_fbtl_posix_component_file_unquery (mca_io_ompio_file_t *file) {
}
int mca_fbtl_posix_module_init (mca_io_ompio_file_t *file) {
#if defined (FBTL_POSIX_HAVE_AIO)
long val = sysconf(_SC_AIO_MAX);
if ( -1 != val ) {
fbtl_posix_max_aio_active_reqs = (int)val;
}
#endif
return OMPI_SUCCESS;
}
@ -107,14 +116,15 @@ bool mca_fbtl_posix_progress ( mca_ompio_request_t *req)
{
bool ret=false;
#if defined (FBTL_POSIX_HAVE_AIO)
int i=0;
int i=0, lcount=0;
mca_fbtl_posix_request_data_t *data=(mca_fbtl_posix_request_data_t *)req->req_data;
for (i=0; i < data->aio_req_count; i++ ) {
for (i=data->aio_first_active_req; i < data->aio_last_active_req; i++ ) {
if ( EINPROGRESS == data->aio_req_status[i] ) {
data->aio_req_status[i] = aio_error ( &data->aio_reqs[i]);
if ( 0 == data->aio_req_status[i]){
data->aio_open_reqs--;
lcount++;
/* assuming right now that aio_return will return
** the number of bytes written/read and not an error code,
** since aio_error should have returned an error in that
@ -135,8 +145,42 @@ bool mca_fbtl_posix_progress ( mca_ompio_request_t *req)
break;
}
}
else {
lcount++;
}
}
#if 0
printf("lcount=%d open_reqs=%d\n", lcount, data->aio_open_reqs );
#endif
if ( (lcount == data->aio_req_chunks) && (0 != data->aio_open_reqs )) {
/* post the next batch of operations */
data->aio_first_active_req = data->aio_last_active_req;
if ( (data->aio_req_count-data->aio_last_active_req) > data->aio_req_chunks ) {
data->aio_last_active_req += data->aio_req_chunks;
}
else {
data->aio_last_active_req = data->aio_req_count;
}
for ( i=data->aio_first_active_req; i< data->aio_last_active_req; i++ ) {
if ( FBTL_POSIX_READ == data->aio_req_type ) {
if (-1 == aio_read(&data->aio_reqs[i])) {
perror("aio_read() error");
return OMPI_ERROR;
}
}
else if ( FBTL_POSIX_WRITE == data->aio_req_type ) {
if (-1 == aio_write(&data->aio_reqs[i])) {
perror("aio_write() error");
return OMPI_ERROR;
}
}
}
#if 0
printf("posting new batch: first=%d last=%d\n", data->aio_first_active_req, data->aio_last_active_req );
#endif
}
if ( 0 == data->aio_open_reqs ) {
/* all pending operations are finished for this request */
req->req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS;

Просмотреть файл

@ -9,7 +9,7 @@
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2008-2014 University of Houston. All rights reserved.
* Copyright (c) 2008-2015 University of Houston. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@ -39,6 +39,8 @@ int mca_fbtl_posix_component_file_unquery (mca_io_ompio_file_t *file);
int mca_fbtl_posix_module_init (mca_io_ompio_file_t *file);
int mca_fbtl_posix_module_finalize (mca_io_ompio_file_t *file);
extern int fbtl_posix_max_aio_active_reqs;
OMPI_MODULE_DECLSPEC extern mca_fbtl_base_component_2_0_0_t mca_fbtl_posix_component;
/*
* ******************************************************************
@ -57,17 +59,25 @@ bool mca_fbtl_posix_progress ( mca_ompio_request_t *req);
void mca_fbtl_posix_request_free ( mca_ompio_request_t *req);
struct mca_fbtl_posix_request_data_t {
int aio_req_count;
int aio_open_reqs;
struct aiocb *aio_reqs;
int *aio_req_status;
ssize_t aio_total_len;
int aio_req_count; /* total number of aio reqs */
int aio_open_reqs; /* number of unfinished reqs */
int aio_req_type; /* read or write */
int aio_req_chunks; /* max. no. of aio reqs that can be posted at once*/
int aio_first_active_req; /* first active posted req */
int aio_last_active_req; /* last currently active poted req */
struct aiocb *aio_reqs; /* pointer array of req structures */
int *aio_req_status; /* array of statuses */
ssize_t aio_total_len; /* total amount of data written */
};
typedef struct mca_fbtl_posix_request_data_t mca_fbtl_posix_request_data_t;
/* Right now statically defined, will become a configure check */
#define FBTL_POSIX_HAVE_AIO 1
/* define constants for AIO requests */
#define FBTL_POSIX_READ 1
#define FBTL_POSIX_WRITE 2
/*
* ******************************************************************
* ************ functions implemented in this module end ************

Просмотреть файл

@ -9,7 +9,7 @@
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2008-2014 University of Houston. All rights reserved.
* Copyright (c) 2008-2015 University of Houston. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@ -47,6 +47,8 @@ ssize_t mca_fbtl_posix_ipreadv (mca_io_ompio_file_t *fh,
data->aio_req_count = fh->f_num_of_io_entries;
data->aio_open_reqs = fh->f_num_of_io_entries;
data->aio_req_type = FBTL_POSIX_READ;
data->aio_req_chunks = fbtl_posix_max_aio_active_reqs;
data->aio_total_len = 0;
data->aio_reqs = (struct aiocb *) malloc (sizeof(struct aiocb) *
fh->f_num_of_io_entries);
@ -70,7 +72,16 @@ ssize_t mca_fbtl_posix_ipreadv (mca_io_ompio_file_t *fh,
data->aio_reqs[i].aio_reqprio = 0;
data->aio_reqs[i].aio_sigevent.sigev_notify = SIGEV_NONE;
data->aio_req_status[i] = EINPROGRESS;
}
data->aio_first_active_req = 0;
if ( data->aio_req_count > data->aio_req_chunks ) {
data->aio_last_active_req = data->aio_req_chunks;
}
else {
data->aio_last_active_req = data->aio_req_count;
}
for (i=0; i < data->aio_last_active_req; i++) {
if (-1 == aio_read(&data->aio_reqs[i])) {
perror("aio_read() error");
return OMPI_ERROR;

Просмотреть файл

@ -9,7 +9,7 @@
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2008-2014 University of Houston. All rights reserved.
* Copyright (c) 2008-2015 University of Houston. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@ -46,6 +46,8 @@ ssize_t mca_fbtl_posix_ipwritev (mca_io_ompio_file_t *fh,
data->aio_req_count = fh->f_num_of_io_entries;
data->aio_open_reqs = fh->f_num_of_io_entries;
data->aio_req_type = FBTL_POSIX_WRITE;
data->aio_req_chunks = fbtl_posix_max_aio_active_reqs;
data->aio_total_len = 0;
data->aio_reqs = (struct aiocb *) malloc (sizeof(struct aiocb) *
fh->f_num_of_io_entries);
@ -69,7 +71,17 @@ ssize_t mca_fbtl_posix_ipwritev (mca_io_ompio_file_t *fh,
data->aio_reqs[i].aio_reqprio = 0;
data->aio_reqs[i].aio_sigevent.sigev_notify = SIGEV_NONE;
data->aio_req_status[i] = EINPROGRESS;
}
data->aio_first_active_req = 0;
if ( data->aio_req_count > data->aio_req_chunks ) {
data->aio_last_active_req = data->aio_req_chunks;
}
else {
data->aio_last_active_req = data->aio_req_count;
}
for (i=0; i < data->aio_last_active_req; i++) {
if (-1 == aio_write(&data->aio_reqs[i])) {
perror("aio_write() error");
return OMPI_ERROR;