ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
7cbc4a4077
openmpi/orte/util/nidmap.c
Ralph Castain ba5498cdc6 Repair the MPI-2 dynamic operations. This includes: 
1. repair of the linear and direct routed modules

2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI

3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature

4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.

Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB

5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.

Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.

6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch

7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch

8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies

9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.

10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon

There is still more cleanup to be done for efficiency, but this at least works.

Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.

Fixes ticket #1256

This commit was SVN r18804.
2008-07-03 17:53:37 +00:00

650 строки
22 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) 2004-2006 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.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "orte_config.h"
#include "orte/types.h"
#include "orte/constants.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <ctype.h>
#include "opal/dss/dss.h"
#include "orte/mca/errmgr/errmgr.h"
#include "orte/util/show_help.h"
#include "orte/util/proc_info.h"
#include "orte/util/name_fns.h"
#include "orte/runtime/orte_globals.h"
#include "orte/util/nidmap.h"
int orte_util_encode_nodemap(opal_byte_object_t *boptr)
{
orte_vpid_t *vpids;
orte_node_t **nodes;
char prefix[ORTE_MAX_NODE_PREFIX], *tmp;
int32_t i, len, firstnode, lastnode, nodenum, num_nodes;
uint8_t command = ORTE_CONTIG_NODE_CMD;
uint8_t num_digs;
uint8_t incdec;
int rc;
char *nodename;
opal_buffer_t buf;
int step;
int32_t *arch;
/* setup a buffer for tmp use */
OBJ_CONSTRUCT(&buf, opal_buffer_t);
/* determine the number of nodes in the global node array */
num_nodes = 0;
nodes = (orte_node_t**)orte_node_pool->addr;
while (num_nodes < orte_node_pool->size &&
NULL != nodes[num_nodes]) {
++num_nodes;
}
/* pack number of nodes */
opal_dss.pack(&buf, &num_nodes, 1, OPAL_INT32);
/* pack the HNP's node name - don't mess with
* trying to encode it - it could be different
*/
/* if we are not keeping FQDN hostnames, abbreviate
* the nodename as required
*/
if (!orte_keep_fqdn_hostnames) {
char *ptr;
nodename = strdup(nodes[0]->name);
if (NULL != (ptr = strchr(nodename, '.'))) {
*ptr = '\0';
}
opal_dss.pack(&buf, &nodename, 1, OPAL_STRING);
free(nodename);
} else {
opal_dss.pack(&buf, &nodes[0]->name, 1, OPAL_STRING);
}
/* see if the cluster is configured with contiguous
* node names and we have more than the HNP
*/
if (orte_contiguous_nodes < num_nodes) {
/* discover the prefix - find first non-alpha character */
len = strlen(nodes[1]->name);
memset(prefix, 0, ORTE_MAX_NODE_PREFIX);
prefix[0] = nodes[1]->name[0]; /* must start with alpha */
for (i=1; i < len; i++) {
if (!isalpha(nodes[1]->name[i])) {
/* found a non-alpha char */
if (!isdigit(nodes[1]->name[i])) {
/* if it is anything but a digit,
* then that's not good
*/
opal_output(0, "%s encode:nidmap Nodename pattern is nonstandard",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
return ORTE_ERROR;
}
/* okay, this defines end of the prefix.
* convert rest of name to an offset
*/
firstnode = strtol(&(nodes[1]->name[i]), NULL, 10);
/* figure out how many digits are in the index */
for (num_digs=0; isdigit(nodes[1]->name[i+num_digs]); num_digs++);
goto PACK;
}
prefix[i] = nodes[1]->name[i];
}
PACK:
/* begin encoding rest of map by indicating that this will
* be a contiguous node map
*/
opal_dss.pack(&buf, &command, 1, OPAL_UINT8);
/* pack the prefix */
tmp = &prefix[0];
opal_dss.pack(&buf, &tmp, 1, OPAL_STRING);
len = strlen(prefix);
/* pack the number of digits in the index */
opal_dss.pack(&buf, &num_digs, 1, OPAL_UINT8);
/* and the starting offset */
opal_dss.pack(&buf, &firstnode, 1, OPAL_INT32);
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s encode:nidmap:contig_nodes prefix %s num_digits %d offset %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), prefix, num_digs, firstnode));
lastnode = strtol(&(nodes[2]->name[i]), NULL, 10);
if ((lastnode - firstnode) < 0) {
/* we are decrementing */
incdec = 0;
opal_dss.pack(&buf, &incdec, 1, OPAL_INT8);
} else {
/* we are incrementing */
incdec = 1;
opal_dss.pack(&buf, &incdec, 1, OPAL_INT8);
}
lastnode = firstnode;
/* cycle through the nodes - pack the starting offset
* and total number of nodes in each contiguous range
*/
for (i=2; i < num_nodes; i++) {
nodenum = strtol(&(nodes[i]->name[len]), NULL, 10);
step = nodenum -lastnode;
if (step < 0) {
/* we are decrementing */
step = lastnode - nodenum;
}
if (step > 1) {
/* have a break - indicate end of range */
opal_dss.pack(&buf, &lastnode, 1, OPAL_INT32);
/* indicate start of new range */
opal_dss.pack(&buf, &nodenum, 1, OPAL_INT32);
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s encode:nidmap:contig_nodes end range %d start next range %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), lastnode, nodenum));
}
lastnode = nodenum;
}
/* pack end of range */
opal_dss.pack(&buf, &lastnode, 1, OPAL_INT32);
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s encode:nidmap:contig_nodes end range %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), lastnode));
/* pack flag end of ranges */
lastnode = -1;
opal_dss.pack(&buf, &lastnode, 1, OPAL_INT32);
} else {
/* if the nodes aren't contiguous, then we need
* to simply pack every nodename individually
*/
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s encode:nidmap non_contig_nodes - packing all names",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* indicate that this will not be a contiguous node map */
command = ORTE_NON_CONTIG_NODE_CMD;
opal_dss.pack(&buf, &command, 1, OPAL_UINT8);
for (i=1; i < num_nodes; i++) {
if (!orte_keep_fqdn_hostnames) {
char *ptr;
nodename = strdup(nodes[i]->name);
if (NULL != (ptr = strchr(nodename, '.'))) {
*ptr = '\0';
}
if (ORTE_SUCCESS != (rc = opal_dss.pack(&buf, &nodename, 1, OPAL_STRING))) {
ORTE_ERROR_LOG(rc);
return rc;
}
free(nodename);
} else {
if (ORTE_SUCCESS != (rc = opal_dss.pack(&buf, &nodes[i]->name, 1, OPAL_STRING))) {
ORTE_ERROR_LOG(rc);
return rc;
}
}
}
}
/* since the daemon vpids may not correspond to the node
* index, we need to also pack the vpid array for all
* daemons. This scenario can happen when the user is
* employing a mapping algo that doesn't use all allocated
* nodes, and sprinkles procs across them in some non-contig
* manner. For example, use of the seq mapper where only
* some nodes are used, and where the usage leaves "holes"
* in the node array, will cause the daemon vpids to not
* match their node array index
*/
/* allocate space for the daemon vpids */
vpids = (orte_vpid_t*)malloc(num_nodes * sizeof(orte_vpid_t));
for (i=0; i < num_nodes; i++) {
if (NULL == nodes[i]->daemon) {
/* some nodes may not have daemons on them */
vpids[i] = ORTE_VPID_INVALID;
continue;
}
vpids[i] = nodes[i]->daemon->name.vpid;
}
opal_dss.pack(&buf, vpids, num_nodes, ORTE_VPID);
free(vpids);
if (OMPI_ENABLE_HETEROGENEOUS_SUPPORT) {
/* check to see if all reported archs are the same */
orte_homogeneous_nodes = true;
for (i=1; i < num_nodes; i++) {
if (nodes[i]->arch != nodes[0]->arch) {
orte_homogeneous_nodes = false;
break;
}
}
if (orte_homogeneous_nodes) {
/* if everything is homo, just set that
* flag - no need to send everything
*/
num_digs = 0;
opal_dss.pack(&buf, &num_digs, 1, OPAL_UINT8);
} else {
/* it isn't homo, so we have to pass the
* archs to the daemons
*/
num_digs = 1;
opal_dss.pack(&buf, &num_digs, 1, OPAL_UINT8);
/* allocate space for the node arch */
arch = (int32_t*)malloc(num_nodes * 4);
/* transfer the data from the nodes */
for (i=0; i < num_nodes; i++) {
arch[i] = nodes[i]->arch;
}
/* pack the values */
opal_dss.pack(&buf, arch, num_nodes, OPAL_INT32);
free(arch);
}
} else {
/* pack a flag indicating that the archs are the same */
num_digs = 0;
opal_dss.pack(&buf, &num_digs, 1, OPAL_UINT8);
}
/* transfer the payload to the byte object */
opal_dss.unload(&buf, (void**)&boptr->bytes, &boptr->size);
OBJ_DESTRUCT(&buf);
return ORTE_SUCCESS;
}
int orte_util_decode_nodemap(opal_byte_object_t *bo, opal_pointer_array_t *nodes)
{
int n, loc, k, diglen, namelen;
char *prefix, digits[10];
int32_t num_nodes, lastnode, endrange, i, num_daemons;
orte_nid_t *node;
orte_vpid_t *vpids;
uint8_t command, num_digs;
orte_nid_t **nd;
uint8_t incdec;
int32_t index, step;
int32_t *arch;
opal_buffer_t buf;
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s decode:nidmap decoding nodemap",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* if there are any entries already in the node array, clear it out */
if (0 < nodes->size) {
/* unfortunately, the opal function "remove_all" doesn't release
* the memory pointed to by the elements in the array, so we need
* to release those first
*/
nd = (orte_nid_t**)nodes->addr;
for (i=0; i < nodes->size && NULL != nd[i]; i++) {
OBJ_RELEASE(nd[i]);
}
/* now use the opal function to reset the internal pointers */
opal_pointer_array_remove_all(nodes);
}
/* xfer the byte object to a buffer for unpacking */
OBJ_CONSTRUCT(&buf, opal_buffer_t);
opal_dss.load(&buf, bo->bytes, bo->size);
/* unpack number of nodes */
n=1;
opal_dss.unpack(&buf, &num_nodes, &n, OPAL_INT32);
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s decode:nidmap decoding %d nodes with %d already loaded",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), num_nodes, nodes->lowest_free));
/* set the size of the nidmap storage so we minimize realloc's */
opal_pointer_array_set_size(nodes, num_nodes);
/* create the struct for the HNP's node */
node = OBJ_NEW(orte_nid_t);
/* the arch defaults to our arch so that non-hetero
* case will yield correct behavior
*/
opal_pointer_array_set_item(nodes, 0, node);
/* unpack the name of the HNP's node */
n=1;
opal_dss.unpack(&buf, &(node->name), &n, OPAL_STRING);
/* unpack flag to see if this is a contiguous node map or not */
n=1;
opal_dss.unpack(&buf, &command, &n, OPAL_UINT8);
if (ORTE_CONTIG_NODE_CMD == command) {
/* unpack the prefix */
n=1;
opal_dss.unpack(&buf, &prefix, &n, OPAL_STRING);
/* the number of digits in the index */
n=1;
opal_dss.unpack(&buf, &num_digs, &n, OPAL_UINT8);
/* and the starting offset */
n=1;
opal_dss.unpack(&buf, &lastnode, &n, OPAL_INT32);
/* unpack increment/decrement flag */
n=1;
opal_dss.unpack(&buf, &incdec, &n, OPAL_INT8);
/* unpack the end of the range */
n=1;
opal_dss.unpack(&buf, &endrange, &n, OPAL_INT32);
/* setup loop params */
if (0 == incdec) {
endrange -= 1;
step = -1;
} else {
endrange += 1;
step = 1;
}
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s decode:nidmap:contig_nodes prefix %s num_digits %d offset %d endrange %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), prefix, num_digs, lastnode, endrange));
namelen = strlen(prefix) + num_digs + 1;
/* cycle through the ranges */
index = 1;
while (1) {
for (i=lastnode; i != endrange; i += step) {
node = OBJ_NEW(orte_nid_t);
/* allocate space for the nodename */
node->name = (char*)malloc(namelen);
memset(node->name, 0, namelen);
loc = snprintf(node->name, namelen, "%s", prefix);
diglen = num_digs - snprintf(digits, 10, "%d", i);
for (k=0; k < diglen && loc < namelen; k++) {
node->name[loc] = '0';
loc++;
}
strncat(node->name, digits, num_digs);
/* the arch defaults to our arch so that non-hetero
* case will yield correct behavior
*/
opal_pointer_array_set_item(nodes, index, node);
index++;
}
/* unpack start of new range */
n=1;
opal_dss.unpack(&buf, &lastnode, &n, OPAL_INT32);
/* if that is -1, then it flags no more ranges */
if (-1 == lastnode) {
goto process_daemons;
}
n=1;
opal_dss.unpack(&buf, &endrange, &n, OPAL_INT32);
if (0 == incdec) {
endrange -= 1;
} else {
endrange += 1;
}
}
} else {
/* not contiguous - just loop over nodes and
* unpack the raw nodename
*/
for (i=1; i < num_nodes; i++) {
node = OBJ_NEW(orte_nid_t);
/* the arch defaults to our arch so that non-hetero
* case will yield correct behavior
*/
opal_pointer_array_set_item(nodes, i, node);
/* unpack the node's name */
n=1;
opal_dss.unpack(&buf, &(node->name), &n, OPAL_STRING);
}
}
process_daemons:
/* unpack the daemon names */
vpids = (orte_vpid_t*)malloc(num_nodes * sizeof(orte_vpid_t));
n=num_nodes;
opal_dss.unpack(&buf, vpids, &n, ORTE_VPID);
/* transfer the data to the nidmap, counting the number of
* daemons in the system
*/
num_daemons = 0;
nd = (orte_nid_t**)nodes->addr;
for (i=0; i < num_nodes; i++) {
nd[i]->daemon = vpids[i];
if (ORTE_VPID_INVALID != vpids[i]) {
++num_daemons;
}
}
free(vpids);
/* if we are a daemon or the HNP, update our num_procs */
if (orte_process_info.hnp || orte_process_info.daemon) {
orte_process_info.num_procs = num_daemons;
}
/* unpack a flag to see if we are in a homogeneous
* scenario - could be that no hetero is supported,
* or could be that things just are homo anyway
*/
n=1;
opal_dss.unpack(&buf, &num_digs, &n, OPAL_UINT8);
if (0 == num_digs) {
/* homo situation */
orte_homogeneous_nodes = true;
} else {
/* hetero situation */
orte_homogeneous_nodes = false;
/* get the archs */
arch = (int32_t*)malloc(num_nodes * 4);
/* unpack the values */
n=num_nodes;
opal_dss.unpack(&buf, arch, &n, OPAL_INT32);
/* transfer the data to the nodes */
nd = (orte_nid_t**)nodes->addr;
for (i=0; i < num_nodes; i++) {
nd[i]->arch = arch[i];
}
free(arch);
}
if (0 < opal_output_get_verbosity(orte_debug_output)) {
nd = (orte_nid_t**)nodes->addr;
for (i=0; i < num_nodes; i++) {
opal_output(0, "%s node[%d].name %s daemon %s arch %0x",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), i,
(NULL == nd[i]) ? "NULL" : nd[i]->name,
ORTE_VPID_PRINT(nd[i]->daemon),
(NULL == nd[i]) ? 0 : nd[i]->arch);
}
}
OBJ_DESTRUCT(&buf);
return ORTE_SUCCESS;
}
int orte_util_encode_pidmap(orte_job_t *jdata, opal_byte_object_t *boptr)
{
int32_t *nodes;
orte_proc_t **procs;
orte_vpid_t i;
int8_t *tmp, flag;
opal_buffer_t buf;
/* setup the working buffer */
OBJ_CONSTRUCT(&buf, opal_buffer_t);
/* pack the number of procs */
opal_dss.pack(&buf, &jdata->num_procs, 1, ORTE_VPID);
/* allocate memory for the nodes */
nodes = (int32_t*)malloc(jdata->num_procs * 4);
/* transfer and pack the node info in one pack */
procs = (orte_proc_t**)jdata->procs->addr;
for (i=0; i < jdata->num_procs; i++) {
nodes[i] = procs[i]->node->index;
}
opal_dss.pack(&buf, nodes, jdata->num_procs, OPAL_INT32);
/* free node storage */
free(nodes);
/* allocate memory for the local_ranks */
tmp = (int8_t*)malloc(jdata->num_procs);
/* transfer and pack them in one pack */
for (i=0; i < jdata->num_procs; i++) {
tmp[i] = procs[i]->local_rank;
}
opal_dss.pack(&buf, tmp, jdata->num_procs, OPAL_UINT8);
/* transfer and pack the node ranks in one pack */
for (i=0; i < jdata->num_procs; i++) {
tmp[i] = procs[i]->node_rank;
}
opal_dss.pack(&buf, tmp, jdata->num_procs, OPAL_UINT8);
/* transfer and pack the app_idx in one pack */
for (i=0; i < jdata->num_procs; i++) {
tmp[i] = procs[i]->app_idx;
}
opal_dss.pack(&buf, tmp, jdata->num_procs, OPAL_INT8);
/* free the storage */
free(tmp);
/* are there cpu_list strings? */
if (jdata->map->cpu_lists) {
flag = (int)true;
opal_dss.pack(&buf, &flag, 1, OPAL_INT8);
for (i=0; i < jdata->num_procs; i++) {
opal_dss.pack(&buf, &procs[i]->slot_list, 1, OPAL_STRING);
}
} else {
flag = (int)false;
opal_dss.pack(&buf, &flag, 1, OPAL_INT8);
}
/* transfer the payload to the byte object */
opal_dss.unload(&buf, (void**)&boptr->bytes, &boptr->size);
OBJ_DESTRUCT(&buf);
return ORTE_SUCCESS;
}
int orte_util_decode_pidmap(opal_byte_object_t *bo, orte_vpid_t *nprocs,
opal_value_array_t *procs, int8_t **app_idx,
char ***slot_str)
{
orte_vpid_t i, num_procs;
orte_pmap_t pmap;
int32_t *nodes;
int8_t *local_rank, *node_rank, *idx;
int8_t flag;
char **slots;
orte_std_cntr_t n;
opal_buffer_t buf;
/* xfer the byte object to a buffer for unpacking */
/* load it into a buffer */
OBJ_CONSTRUCT(&buf, opal_buffer_t);
opal_dss.load(&buf, bo->bytes, bo->size);
/* unpack the number of procs */
n=1;
opal_dss.unpack(&buf, &num_procs, &n, ORTE_VPID);
*nprocs = num_procs;
/* allocate memory for the procs array */
opal_value_array_set_size(procs, num_procs);
/* allocate memory for the node info */
nodes = (int32_t*)malloc(num_procs * 4);
/* unpack it in one shot */
n=num_procs;
opal_dss.unpack(&buf, nodes, &n, OPAL_INT32);
/* allocate memory for local ranks */
local_rank = (int8_t*)malloc(num_procs);
/* unpack them in one shot */
n=num_procs;
opal_dss.unpack(&buf, local_rank, &n, OPAL_UINT8);
/* allocate memory for node ranks */
node_rank = (int8_t*)malloc(num_procs);
/* unpack node ranks in one shot */
n=num_procs;
opal_dss.unpack(&buf, node_rank, &n, OPAL_UINT8);
/* store the data */
for (i=0; i < num_procs; i++) {
pmap.node = nodes[i];
pmap.local_rank = local_rank[i];
pmap.node_rank = node_rank[i];
opal_value_array_set_item(procs, i, &pmap);
}
/* release data */
free(nodes);
free(local_rank);
free(node_rank);
/* only daemons/HNPs need the rest of the data, so if
* we aren't one of those, we are done!
*/
if (!orte_process_info.hnp &&
!orte_process_info.daemon) {
OBJ_DESTRUCT(&buf);
return ORTE_SUCCESS;
}
/* allocate memory for app_idx */
idx = (int8_t*)malloc(num_procs);
/* unpack app_idx in one shot */
n=num_procs;
opal_dss.unpack(&buf, idx, &n, OPAL_INT8);
/* hand the array back to the caller */
*app_idx = idx;
/* unpack flag to indicate if slot_strings are present */
n=1;
opal_dss.unpack(&buf, &flag, &n, OPAL_INT8);
if (flag) {
/* allocate space */
slots = (char**)malloc(num_procs * sizeof(char*));
for (i=0; i < num_procs; i++) {
n=1;
opal_dss.unpack(&buf, &slots[i], &n, OPAL_STRING);
}
*slot_str = slots;
}
OBJ_DESTRUCT(&buf);
return ORTE_SUCCESS;
}
Показать git blame Копировать постоянную ссылку