1
1
openmpi/ompi/mca/bcol/basesmuma/bcol_basesmuma_smcm.c

455 строки
16 KiB
C
Исходник Обычный вид История

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
*
* Copyright (c) 2009-2012 Oak Ridge National Laboratory. All rights reserved.
* Copyright (c) 2009-2012 Mellanox Technologies. All rights reserved.
* Copyright (c) 2012-2014 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2014 Intel, Inc. All rights reserved.
* Copyright (c) 2014 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include <stdlib.h>
#include <stdio.h>
#include <fcntl.h>
#include <errno.h>
#ifdef HAVE_STRINGS_H
#include <strings.h>
#endif
#include "ompi/proc/proc.h"
#include "ompi/patterns/comm/coll_ops.h"
#include "opal/align.h"
#include "opal/dss/dss.h"
#include "opal/util/error.h"
#include "opal/util/output.h"
#include "opal/class/opal_list.h"
#include "opal/class/opal_hash_table.h"
#include "bcol_basesmuma.h"
#define SM_BACKING_FILE_NAME_MAX_LEN 256
static bcol_basesmuma_smcm_mmap_t * bcol_basesmuma_smcm_reg_mmap(void *in_ptr, int fd, size_t length,
size_t addr_offset, size_t alignment,
char *file_name);
struct file_info_t {
uint32_t vpid;
uint32_t jobid;
uint64_t file_size;
uint64_t size_ctl_structure;
uint64_t data_seg_alignment;
char file_name[SM_BACKING_FILE_NAME_MAX_LEN];
};
/* need to allocate space for the peer */
static void bcol_basesmuma_smcm_proc_item_t_construct (bcol_basesmuma_smcm_proc_item_t * item)
{
memset ((char *) item + sizeof (item->item), 0, sizeof (*item) - sizeof (item->item));
}
/* need to free the space for the peer */
static void bcol_basesmuma_smcm_proc_item_t_destruct (bcol_basesmuma_smcm_proc_item_t * item)
{
if (item->sm_mmap) {
OBJ_RELEASE(item->sm_mmap);
}
if (item->sm_file.file_name) {
free (item->sm_file.file_name);
item->sm_file.file_name = NULL;
}
}
OBJ_CLASS_INSTANCE(bcol_basesmuma_smcm_proc_item_t,
opal_list_item_t,
bcol_basesmuma_smcm_proc_item_t_construct,
bcol_basesmuma_smcm_proc_item_t_destruct);
static void bcol_basesmuma_smcm_mmap_construct (bcol_basesmuma_smcm_mmap_t *smcm_mmap)
{
memset ((char *) smcm_mmap + sizeof (smcm_mmap->super), 0, sizeof (*smcm_mmap) - sizeof (smcm_mmap->super));
}
static void bcol_basesmuma_smcm_mmap_destruct (bcol_basesmuma_smcm_mmap_t *smcm_mmap)
{
if (smcm_mmap->map_seg) {
munmap (smcm_mmap->map_seg, smcm_mmap->map_size);
smcm_mmap->map_seg = NULL;
}
if (smcm_mmap->map_path) {
free (smcm_mmap->map_path);
smcm_mmap->map_path = NULL;
}
}
OBJ_CLASS_INSTANCE(bcol_basesmuma_smcm_mmap_t, opal_list_item_t,
bcol_basesmuma_smcm_mmap_construct,
bcol_basesmuma_smcm_mmap_destruct);
/* smcm_allgather_connection:
This function is called when a shared memory subgroup wants to establish shared memory "connections" among
a group of processes.
This function DOES NOT create any shared memory backing files, it only mmaps already existing files. Shared
memory files are created by the shared memory registration function
-----------------------------------------------------------------------------------------------------------
Input params:
- sbgp module The subgrouping module contains the list of ranks to wire up.
- peer_list An opal list containing a list of bcol_basesmuma_smcm_proc_item_t types. This
contains a list of peers whose shared memory files I have already mapped.
Upon completion of the allgather exchange with all members of the group and depending on the
value of "map_all", my peers' shared memory files are mapped into my local virtual memory
space, with all pertinent information being stored in an bcol_basesmuma_smcm_proc_item_t which is
subsequently appended onto the "peer_list".
- comm The ompi_communicator_t communicator.
- input A data struct that caches the information about my shared memory file.
- map_all Bool that determines whether or not to go ahead and map the files from all of the peers
defined in the sbgp-ing module. If map_all == true, then go ahead and mmap all of the files
obtained in the exchange and append the information to the "peer_list". If map_all == false
then make a check and only mmap those peers' files whose vpid/jobid/filename combination do
not already exist in the "peer_list". Once mapping is completed, append this peer's information
to the "peer_list".
-----------------------------------------------------------------------------------------------------------
*
*/
int bcol_basesmuma_smcm_allgather_connection(
mca_bcol_basesmuma_module_t *sm_bcol_module,
mca_sbgp_base_module_t *module,
opal_list_t *peer_list,
bcol_basesmuma_smcm_proc_item_t ***back_files,
ompi_communicator_t *comm,
bcol_basesmuma_smcm_file_t input,
char *base_fname,
bool map_all)
{
/* define local variables */
int rc, i, fd;
ptrdiff_t mem_offset;
ompi_proc_t *proc_temp, *my_id;
bcol_basesmuma_smcm_proc_item_t *temp;
bcol_basesmuma_smcm_proc_item_t *item_ptr;
bcol_basesmuma_smcm_proc_item_t **backing_files;
struct file_info_t local_file;
struct file_info_t *all_files=NULL;
/* sanity check */
if (strlen(input.file_name) > SM_BACKING_FILE_NAME_MAX_LEN-1) {
opal_output (ompi_bcol_base_framework.framework_output, "backing file name too long: %s len :: %d",
input.file_name, (int) strlen(input.file_name));
return OMPI_ERR_BAD_PARAM;
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
}
backing_files = (bcol_basesmuma_smcm_proc_item_t **)
calloc(module->group_size, sizeof(bcol_basesmuma_smcm_proc_item_t *));
if (!backing_files) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* FIXME *back_files might have been already allocated
* so free it in order to avoid a memory leak */
if (NULL != *back_files) {
free (*back_files);
}
*back_files = backing_files;
my_id = ompi_proc_local();
/* Phase One:
gather a list of processes that will participate in the allgather - I'm
preparing this list from the sbgp-ing module that was passed into the function */
/* fill in local file information */
local_file.vpid=my_id->proc_name.vpid;
local_file.jobid=my_id->proc_name.jobid;
local_file.file_size=input.size;
local_file.size_ctl_structure=input.size_ctl_structure;
local_file.data_seg_alignment=input.data_seg_alignment;
strcpy (local_file.file_name, input.file_name);
/* will exchange this data type as a string of characters -
* this routine is first called before MPI_init() completes
* and before error handling is setup, so can't use the
* MPI data types to send this data */
all_files = (struct file_info_t *) calloc(module->group_size,
sizeof (struct file_info_t));
if (!all_files) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* exchange data */
rc = comm_allgather_pml(&local_file,all_files,sizeof(struct file_info_t), MPI_CHAR,
sm_bcol_module->super.sbgp_partner_module->my_index,
sm_bcol_module->super.sbgp_partner_module->group_size,
sm_bcol_module->super.sbgp_partner_module->group_list,
sm_bcol_module->super.sbgp_partner_module->group_comm);
if( OMPI_SUCCESS != rc ) {
opal_output (ompi_bcol_base_framework.framework_output, "failed in comm_allgather_pml. Error code: %d", rc);
goto Error;
}
/* Phase four:
loop through the receive buffer, unpack the data recieved from remote peers */
for (i = 0; i < module->group_size; i++) {
struct file_info_t *rem_file = all_files + i;
/* check if this is my index or if the file is already mapped (set above). ther
* is no reason to look through the peer list again because no two members of
* the group will have the same vpid/jobid pair. ignore this previously found
* mapping if map_all was requested (NTH: not sure why exactly since we re-map
* and already mapped file) */
if (sm_bcol_module->super.sbgp_partner_module->my_index == i) {
continue;
}
proc_temp = ompi_comm_peer_lookup(comm,module->group_list[i]);
OPAL_LIST_FOREACH(item_ptr, peer_list, bcol_basesmuma_smcm_proc_item_t) {
/* if the vpid/jobid/filename combination already exists in the list,
then do not map this peer's file --- because you already have */
if (proc_temp->proc_name.vpid == item_ptr->peer.vpid &&
proc_temp->proc_name.jobid == item_ptr->peer.jobid &&
0 == strcmp (item_ptr->sm_file.file_name, rem_file->file_name)) {
++item_ptr->refcnt;
/* record file data */
backing_files[i] = item_ptr;
break;
}
}
if (!map_all && backing_files[i]) {
continue;
}
temp = OBJ_NEW(bcol_basesmuma_smcm_proc_item_t);
if (!temp) {
rc = OMPI_ERR_OUT_OF_RESOURCE;
goto Error;
}
temp->peer.vpid = rem_file->vpid;
temp->peer.jobid = rem_file->jobid;
temp->sm_file.file_name = strdup (rem_file->file_name);
if (!temp->sm_file.file_name) {
rc = OMPI_ERR_OUT_OF_RESOURCE;
OBJ_RELEASE(temp);
goto Error;
}
temp->sm_file.size = (size_t) rem_file->file_size;
temp->sm_file.mpool_size = (size_t) rem_file->file_size;
temp->sm_file.size_ctl_structure = (size_t) rem_file->size_ctl_structure;
temp->sm_file.data_seg_alignment = (size_t) rem_file->data_seg_alignment;
temp->refcnt = 1;
/* Phase Five:
If map_all == true, then we map every peer's file
else we check to see if I have already mapped this
vpid/jobid/filename combination and if I have, then
I do not mmap this peer's file.
*
*/
fd = open(temp->sm_file.file_name, O_RDWR, 0600);
if (0 > fd) {
opal_output (ompi_bcol_base_framework.framework_output, "SMCM Allgather failed to open sm backing file %s. errno = %d",
temp->sm_file.file_name, errno);
rc = OMPI_ERROR;
goto Error;
}
/* map the file */
temp->sm_mmap = bcol_basesmuma_smcm_reg_mmap (NULL, fd, temp->sm_file.size,
temp->sm_file.size_ctl_structure,
temp->sm_file.data_seg_alignment,
temp->sm_file.file_name);
close (fd);
if (NULL == temp->sm_mmap) {
opal_output (ompi_bcol_base_framework.framework_output, "mmapping failed to map remote peer's file");
OBJ_RELEASE(temp);
rc = OMPI_ERROR;
goto Error;
}
/* compute memory offset */
mem_offset = (ptrdiff_t) temp->sm_mmap->data_addr -
(ptrdiff_t) temp->sm_mmap->map_seg;
temp->sm_mmap->map_seg->seg_offset = mem_offset;
temp->sm_mmap->map_seg->seg_size = temp->sm_file.size - mem_offset;
/* more stuff to follow */
/* append this peer's info, including shared memory map addr, onto the
peer_list */
/* record file data */
backing_files[i] = (bcol_basesmuma_smcm_proc_item_t *) temp;
opal_list_append(peer_list, (opal_list_item_t*) temp);
}
rc = OMPI_SUCCESS;
Error:
/* error clean-up and return */
if (NULL != all_files) {
free(all_files);
}
return rc;
}
int bcol_basesmuma_smcm_release_connections (mca_bcol_basesmuma_module_t *sm_bcol_module,
mca_sbgp_base_module_t *sbgp_module, opal_list_t *peer_list,
bcol_basesmuma_smcm_proc_item_t ***back_files)
{
bcol_basesmuma_smcm_proc_item_t **smcm_procs = *back_files;
for (int i = 0 ; i < sbgp_module->group_size ; ++i) {
if (smcm_procs[i] && 0 == --smcm_procs[i]->refcnt) {
opal_list_remove_item (peer_list, (opal_list_item_t *) smcm_procs[i]);
OBJ_RELEASE(smcm_procs[i]);
}
}
free (smcm_procs);
*back_files = NULL;
return OMPI_SUCCESS;
}
/*
* mmap the specified file as a shared file. No information exchange with other
* processes takes place within this routine.
* This function assumes that the memory has already been allocated, and only the
* mmap needs to be done.
*/
bcol_basesmuma_smcm_mmap_t *bcol_basesmuma_smcm_mem_reg(void *in_ptr,
size_t length,
size_t alignment,
char* file_name)
{
/* local variables */
int fd = -1;
bcol_basesmuma_smcm_mmap_t *map = NULL;
int rc;
/* if pointer is not allocated - return error. We have no clue how the user will allocate or
* free this memory.
*/
/* open the shared memory backing file */
fd = open(file_name, O_CREAT|O_RDWR,0600);
if (fd < 0) {
opal_output (ompi_bcol_base_framework.framework_output, "basesmuma shared memory allocation open failed with errno: %d",
errno);
} else if (0 != ftruncate(fd,length)) {
opal_output (ompi_bcol_base_framework.framework_output, "basesmuma shared memory allocation ftruncate failed with errno: %d",
errno);
} else {
/* ensure there is enough space for the backing store */
rc = ftruncate (fd, length);
if (0 > rc) {
opal_output (ompi_bcol_base_framework.framework_output, "failed to truncate the file to be mapped. errno: %d", errno);
return NULL;
}
map = bcol_basesmuma_smcm_reg_mmap(in_ptr, fd, length, 0, alignment, file_name);
if (NULL == map) {
return NULL;
}
}
/* no longer need this file descriptor. close it */
close (fd);
/* takes us to the top of the control structure */
return map;
}
static bcol_basesmuma_smcm_mmap_t * bcol_basesmuma_smcm_reg_mmap(void *in_ptr, int fd, size_t length,
size_t addr_offset, size_t alignment,
char *file_name)
{
/* local variables */
bcol_basesmuma_smcm_mmap_t *map;
bcol_basesmuma_smcm_file_header_t *seg;
unsigned char* myaddr = NULL;
int flags = MAP_SHARED;
/* set up the map object */
map = OBJ_NEW(bcol_basesmuma_smcm_mmap_t);
if (OPAL_UNLIKELY(NULL == map)) {
return NULL;
}
/* map the file and initialize the segment state */
if (NULL != in_ptr) {
flags |= MAP_FIXED;
}
seg = (bcol_basesmuma_smcm_file_header_t *)
mmap(in_ptr, length, PROT_READ|PROT_WRITE, flags, fd, 0);
if((void*)-1 == seg) {
OBJ_RELEASE(map);
return NULL;
}
map->map_path = strdup (file_name);
/* the first entry in the file is the control structure. the first entry
in the control structure is an mca_common_sm_file_header_t element */
map->map_seg = seg;
myaddr = (unsigned char *) seg + addr_offset;
/* if we have a data segment (i.e. if 0 != data_seg_alignement) */
if (alignment) {
myaddr = OPAL_ALIGN_PTR(myaddr, alignment, unsigned char*);
/* is addr past the end of the file? */
if ((unsigned char *) seg+length < myaddr) {
opal_output (ompi_bcol_base_framework.framework_output, "mca_bcol_basesmuma_sm_alloc_mmap: memory region too small len %lu add %p",
(unsigned long) length, myaddr);
OBJ_RELEASE(map);
munmap (seg, length);
return NULL;
}
}
map->data_addr = (unsigned char*) myaddr;
map->map_addr = (unsigned char*) seg;
map->map_size = length;
return map;
}