1
1
openmpi/ompi/mca/btl/udapl/btl_udapl_proc.c

311 строки
11 KiB
C
Исходник Обычный вид История

/*
* Copyright (c) 2004-2006 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2011 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) 2006 Sandia National Laboratories. All rights
* reserved.
* Copyright (c) 2008 Sun Microsystems, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "ompi/runtime/ompi_module_exchange.h"
#include "opal/util/net.h"
#include "btl_udapl.h"
#include "btl_udapl_endpoint.h"
#include "btl_udapl_proc.h"
static void mca_btl_udapl_proc_construct(mca_btl_udapl_proc_t* proc);
static void mca_btl_udapl_proc_destruct(mca_btl_udapl_proc_t* proc);
OBJ_CLASS_INSTANCE(mca_btl_udapl_proc_t,
opal_list_item_t, mca_btl_udapl_proc_construct,
mca_btl_udapl_proc_destruct);
void mca_btl_udapl_proc_construct(mca_btl_udapl_proc_t* udapl_proc)
{
udapl_proc->proc_ompi = 0;
udapl_proc->proc_addr_count = 0;
udapl_proc->proc_endpoints = 0;
udapl_proc->proc_endpoint_count = 0;
OBJ_CONSTRUCT(&udapl_proc->proc_lock, opal_mutex_t);
/* add to list of all proc instance */
OPAL_THREAD_LOCK(&mca_btl_udapl_component.udapl_lock);
opal_list_append(&mca_btl_udapl_component.udapl_procs, &udapl_proc->super);
OPAL_THREAD_UNLOCK(&mca_btl_udapl_component.udapl_lock);
}
/*
* Cleanup uDAPL proc instance
*/
void mca_btl_udapl_proc_destruct(mca_btl_udapl_proc_t* udapl_proc)
{
/* remove from list of all proc instances */
OPAL_THREAD_LOCK(&mca_btl_udapl_component.udapl_lock);
opal_list_remove_item(&mca_btl_udapl_component.udapl_procs, &udapl_proc->super);
OPAL_THREAD_UNLOCK(&mca_btl_udapl_component.udapl_lock);
/* release resources */
if(NULL != udapl_proc->proc_endpoints) {
free(udapl_proc->proc_endpoints);
}
OBJ_DESTRUCT(&udapl_proc->proc_lock);
}
/*
* Look for an existing uDAPL process instances based on the associated
* ompi_proc_t instance.
*/
static mca_btl_udapl_proc_t* mca_btl_udapl_proc_lookup_ompi(ompi_proc_t* ompi_proc)
{
mca_btl_udapl_proc_t* udapl_proc;
OPAL_THREAD_LOCK(&mca_btl_udapl_component.udapl_lock);
for(udapl_proc = (mca_btl_udapl_proc_t*)
opal_list_get_first(&mca_btl_udapl_component.udapl_procs);
udapl_proc != (mca_btl_udapl_proc_t*)
opal_list_get_end(&mca_btl_udapl_component.udapl_procs);
udapl_proc = (mca_btl_udapl_proc_t*)opal_list_get_next(udapl_proc)) {
if(udapl_proc->proc_ompi == ompi_proc) {
OPAL_THREAD_UNLOCK(&mca_btl_udapl_component.udapl_lock);
return udapl_proc;
}
}
OPAL_THREAD_UNLOCK(&mca_btl_udapl_component.udapl_lock);
return NULL;
}
/*
* Create a uDAPL process structure. There is a one-to-one correspondence
* between a ompi_proc_t and a mca_btl_udapl_proc_t instance. We cache
* additional data (specifically the list of mca_btl_udapl_endpoint_t instances,
* and published addresses) associated w/ a given destination on this
* datastructure.
*/
mca_btl_udapl_proc_t* mca_btl_udapl_proc_create(ompi_proc_t* ompi_proc)
{
mca_btl_udapl_proc_t* udapl_proc = NULL;
size_t size;
int rc;
/* Check if we have already created a uDAPL proc
* structure for this ompi process */
udapl_proc = mca_btl_udapl_proc_lookup_ompi(ompi_proc);
if(udapl_proc != NULL) {
return udapl_proc;
}
/* create a new udapl proc out of the ompi_proc ... */
udapl_proc = OBJ_NEW(mca_btl_udapl_proc_t);
udapl_proc->proc_endpoint_count = 0;
udapl_proc->proc_ompi = ompi_proc;
/* query for the peer address info */
rc = ompi_modex_recv(
&mca_btl_udapl_component.super.btl_version,
ompi_proc,
(void*)&udapl_proc->proc_addrs,
&size);
if(OMPI_SUCCESS != rc) {
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_CRITICAL,
("ompi_modex_recv failed for peer %s",
OMPI_NAME_PRINT(&ompi_proc->proc_name)));
OBJ_RELEASE(udapl_proc);
return NULL;
}
if((size % sizeof(mca_btl_udapl_addr_t)) != 0) {
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_CRITICAL,
("invalid udapl address for peer %s",
OMPI_NAME_PRINT(&ompi_proc->proc_name)));
OBJ_RELEASE(udapl_proc);
return NULL;
}
udapl_proc->proc_addr_count = size/sizeof(mca_btl_udapl_addr_t);
if (0 == udapl_proc->proc_addr_count) {
udapl_proc->proc_endpoints = NULL;
} else {
udapl_proc->proc_endpoints = (mca_btl_base_endpoint_t**)
malloc(udapl_proc->proc_addr_count * sizeof(mca_btl_base_endpoint_t*));
}
if(NULL == udapl_proc->proc_endpoints) {
OBJ_RELEASE(udapl_proc);
return NULL;
}
return udapl_proc;
}
/*
* Find an address on the peer_process which matches stated criteria
* to the udapl btl module address information. Return in peer_addr_idx
* the index to the peer_process address that matches the btl module
* address. Where match criteria is:
* - the address in not already in use
* - compare addresses using netmask, the netmask value can be modified with
* "--mca btl_udapl_if_mask"
*
* Note: since this is called from mca_btl_udapl_proc_insert() it
* is assumed that the process lock is locked when entered.
*
* @param udapl_btl (IN) BTL module
* @param peer_process (IN) BTL peer process
* @param peer_addr_idx(IN/OUT) Index of address on peer_process
* which matches the udapl_btl address data.
* On success should be >= 0.
* @return OMPI_SUCCESS or error status on failure
*/
static int mca_btl_udapl_proc_address_match(
mca_btl_udapl_module_t* udapl_btl,
mca_btl_udapl_proc_t* peer_proc,
int* peer_addr_idx)
{
int i;
struct sockaddr *saddr;
struct sockaddr_in *btl_addr;
struct sockaddr_in *peer_addr;
char btl_addr_string[INET_ADDRSTRLEN];
char peer_addr_string[INET_ADDRSTRLEN];
*peer_addr_idx = MCA_BTL_UDAPL_INVALID_PEER_ADDR_IDX;
/* use generic address to find address family */
saddr = (struct sockaddr *)&(udapl_btl->udapl_addr.addr);
if (saddr->sa_family == AF_INET) {
btl_addr = (struct sockaddr_in *)saddr;
/* Loop thru peer process addresses looking for match.
* Match criteria:
* - address should not be "inuse"
* - both udapl btl module and peer address should be on
* the same subnet (compare with if_mask value)
*/
for(i = 0; i < (int) peer_proc->proc_addr_count; i++) {
peer_addr =
(struct sockaddr_in *)&(peer_proc->proc_addrs[i].addr);
if (VERBOSE_INFORM <=
mca_btl_udapl_component.udapl_verbosity) {
/* retrieve udapl btl and peer address string for reporting */
inet_ntop(AF_INET, (void *) &btl_addr->sin_addr,
btl_addr_string, INET_ADDRSTRLEN);
inet_ntop(AF_INET, (void *) &peer_addr->sin_addr,
peer_addr_string, INET_ADDRSTRLEN);
}
if ((false == peer_proc->proc_addrs[i].inuse) &&
(opal_net_samenetwork((struct sockaddr *)btl_addr,
(struct sockaddr *)peer_addr, udapl_btl->udapl_if_mask))) {
/* capture index of remote address where match found */
*peer_addr_idx = i;
/* mark this address as now being used */
peer_proc->proc_addrs[i].inuse = true;
/* report what address was found to match */
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_INFORM,
("uDAPL BTL module(%s) matched %s",
btl_addr_string, peer_addr_string));
break;
} else {
/* peer address already used by another udapl btl
* module or netmask check not successful so skip
*/
BTL_UDAPL_VERBOSE_OUTPUT(VERBOSE_INFORM,
("uDAPL BTL module(%s) either skipped because it "
"is already in use or match criteria not successful "
"for peer address %s",
btl_addr_string, peer_addr_string));
}
}
} else {
/* current uDAPL BTL only supports IPv4 */
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP,
("help-mpi-btl-udapl.txt", "IPv4 only",
true, ompi_process_info.nodename));
return OMPI_ERROR;
}
if (MCA_BTL_UDAPL_INVALID_PEER_ADDR_IDX == *peer_addr_idx) {
BTL_UDAPL_VERBOSE_HELP(VERBOSE_SHOW_HELP,
("help-mpi-btl-udapl.txt", "no network match",
true, btl_addr_string, ompi_process_info.nodename,
When we direct launch an application, we rely on PMI for wireup support. In doing so, we lose the de facto data compression we get from the ORTE modex since we no longer get all the wireup info from every proc in a single blob. Instead, we have to iterate over all the procs, calling PMI_KVS_get for every value we require. This creates a really bad scaling behavior. Users have found a nearly 20% launch time differential between mpirun and PMI, with PMI being the slower method. Some of the problem is attributable to poor exchange algorithms in RM's like Slurm and Alps, but we make things worse by calling "get" so many times. Nathan (with a tad advice from me) has attempted to alleviate this problem by reducing the number of "get" calls. This required the following changes: * upon first request for data, have the OPAL db pmi component fetch and decode *all* the info from a given remote proc. It turned out we weren't caching the info, so we would continually request it and only decode the piece we needed for the immediate request. We now decode all the info and push it into the db hash component for local storage - and then all subsequent retrievals are fulfilled locally * reduced the amount of data by eliminating the exchange of the OMPI_ARCH value if heterogeneity is not enabled. This was used solely as a check so we would error out if the system wasn't actually homogeneous, which was fine when we thought there was no cost in doing the check. Unfortunately, at large scale and with direct launch, there is a non-zero cost of making this test. We are open to finding a compromise (perhaps turning the test off if requested?), if people feel strongly about performing the test * reduced the amount of RTE data being automatically fetched, and fetched the rest only upon request. In particular, we no longer immediately fetch the hostname (which is only used for error reporting), but instead get it when needed. Likewise for the RML uri as that info is only required for some (not all) environments. In addition, we no longer fetch the locality unless required, relying instead on the PMI clique info to tell us who is on our local node (if additional info is required, the fetch is performed when a modex_recv is issued). Again, all this only impacts direct launch - all the info is provided when launched via mpirun as there is no added cost to getting it Barring objections, we may move this (plus any required other pieces) to the 1.7 branch once it soaks for an appropriate time. This commit was SVN r29040.
2013-08-17 00:49:18 +00:00
ompi_proc_get_hostname(peer_proc->proc_ompi)));
return OMPI_ERR_OUT_OF_RESOURCE;
}
return OMPI_SUCCESS;
}
/*
* Note that this routine must be called with the lock on the process
* already held. Insert a btl instance into the proc array and assign
* it an address.
*/
int mca_btl_udapl_proc_insert(
mca_btl_udapl_proc_t* udapl_proc,
mca_btl_base_endpoint_t* udapl_endpoint)
{
int peer_address_idx;
mca_btl_udapl_module_t* udapl_btl = udapl_endpoint->endpoint_btl;
/* Check so as not to create more endpoints than addresses.
* Example: If one node has 3 btl modules and another only has 2,
* this check prevents the node with 3 btl modules from
* overloading the other, i.e. only 2 possible connections will
* be possible.
*/
if (udapl_proc->proc_endpoint_count > udapl_proc->proc_addr_count)
return OMPI_ERR_OUT_OF_RESOURCE;
/* Find an endpoint on the udapl process of interest that matches
* the endpoint information of the current udapl btl module
*/
if (OMPI_SUCCESS !=
mca_btl_udapl_proc_address_match(udapl_btl, udapl_proc,
&peer_address_idx)) {
/* no address on peer proc met criteria */
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* insert into endpoint array */
udapl_endpoint->endpoint_proc = udapl_proc;
udapl_endpoint->endpoint_addr =
udapl_proc->proc_addrs[peer_address_idx];
udapl_proc->proc_endpoints[udapl_proc->proc_endpoint_count] = udapl_endpoint;
udapl_proc->proc_endpoint_count++;
return OMPI_SUCCESS;
}