From 235d6029b5997eb0fa61aa1b8deceb630b9d7e6b Mon Sep 17 00:00:00 2001
From: Moe Jette <jette1@llnl.gov>
Date: Wed, 12 Apr 2006 16:34:50 +0000
Subject: [PATCH] svn merge -r7768:7773
https://eris.llnl.gov/svn/slurm/branches/slurm-1.0
---
NEWS | 4 ++++
doc/html/configurator.html | 22 +++++++++++++++++++++-
src/common/env.c | 4 ++--
src/plugins/mpi/mpichgm/mpichgm.c | 23 ++++++++++++-----------
src/srun/io.c | 2 +-
5 files changed, 40 insertions(+), 15 deletions(-)
diff --git a/NEWS b/NEWS
index b1dc07ab0e1..70a072a060e 100644
--- a/NEWS
+++ b/NEWS
@@ -79,6 +79,10 @@ documents those changes that are of interest to users and admins.
-- Make select/cons_res interoperate with mpi/lam plugin for task counts.
-- Fix race condition where srun could seg-fault due to use of logging functions
within pthread after calling log_fini.
+ -- Code changes for clean build with gcc 2.96 (gcc_2_96.patch, Takao Hatazaki, HP).
+ -- Add CacheGroups configuration support in configurator.html (configurator.patch,
+ Takao Hatazaki, HP).
+ -- Fix bug preventing use of mpich-gm plugin (mpichgm.patch, Takao Hatazaki, HP).
* Changes in SLURM 1.0.9
========================
diff --git a/doc/html/configurator.html b/doc/html/configurator.html
index 8aa166aaaee..729bff00e1e 100644
--- a/doc/html/configurator.html
+++ b/doc/html/configurator.html
@@ -77,7 +77,7 @@ function displayfile()
"SlurmdPidFile=" + document.config.slurmd_pid_file.value + "<br>" +
"ProctrackType=proctrack/" + get_radio_value(document.config.proctrack_type) + "<br>" +
"#PluginDir= <br>" +
- "#CacheGroups= <br>" +
+ "CacheGroups=" + get_radio_value(document.config.cache_groups) + "<br>" +
"#FirstJobId= <br>" +
"ReturnToService=" + get_radio_value(document.config.return_to_service) + "<br>" +
"#MaxJobCount= <br>" +
@@ -215,6 +215,26 @@ purposes any user name can be used.
<input type="text" name="slurm_user" value="slurm"> <B>SlurmUser</B>
<P>
+<H2>Group ID Caching</H2>
+
+If you have a slow NIS environment, big parallel jobs take a long time
+to start up (and may eventually time-out) because the NIS server(s)
+may not be able to quickly respond to simultaneous requests from
+multiple slurmd's. You can instruct slurmd to cache /etc/groups
+entries to prevent this from happening by setting
+<B>CacheGroups</B>=1". Reconfiguring ("scontrol reconfig") with
+<B>CacheGroups</B>=0 will cause slurmd to purge the cache. Select one
+value for <B>CacheGroups</B>:<BR>
+<input type="radio" name="cache_groups" value="0" checked>
+<B>0</B>: for normal environment.<BR>
+<input type="radio" name="cache_groups" value="1">
+<B>1</B>: for slow NIS environment.
+<P>
+WARNING: The group ID cache does not try to keep itself in sync with
+the system. You MUST run "scontrol reconfig" to update the cache
+after making any changes to system password or group databases.
+<P>
+
<H2>SLURM Port Numbers</H2>
The SLURM controller (slurmctld) requires a unique port for communcations
as do the SLURM compute node deamonds (slurmd). If not set, slurm ports
diff --git a/src/common/env.c b/src/common/env.c
index 6e14704c738..97ca68e7ca3 100644
--- a/src/common/env.c
+++ b/src/common/env.c
@@ -278,8 +278,8 @@ int setup_env(env_t *env)
}
if (env->cpu_bind_type) {
- unsetenvp(env->env, "SLURM_CPU_BIND"); /* don't propagate SLURM_CPU_BIND */
int setstat = 0;
+ unsetenvp(env->env, "SLURM_CPU_BIND"); /* don't propagate SLURM_CPU_BIND */
if (env->cpu_bind_type & CPU_BIND_VERBOSE) {
setstat |= setenvf(&env->env, "SLURM_CPU_BIND_VERBOSE", "verbose");
} else {
@@ -320,9 +320,9 @@ int setup_env(env_t *env)
rc = SLURM_FAILURE;
}
} else {
+ int setstat = 0;
unsetenvp(env->env, "SLURM_CPU_BIND"); /* don't propagate SLURM_CPU_BIND */
/* set SLURM_CPU_BIND_* env vars to defaults */
- int setstat = 0;
setstat |= setenvf(&env->env, "SLURM_CPU_BIND_VERBOSE", "quiet");
setstat |= setenvf(&env->env, "SLURM_CPU_BIND_TYPE", "");
setstat |= setenvf(&env->env, "SLURM_CPU_BIND_LIST", "");
diff --git a/src/plugins/mpi/mpichgm/mpichgm.c b/src/plugins/mpi/mpichgm/mpichgm.c
index 3d1052054f4..3c6cf86391b 100644
--- a/src/plugins/mpi/mpichgm/mpichgm.c
+++ b/src/plugins/mpi/mpichgm/mpichgm.c
@@ -60,13 +60,10 @@ typedef struct {
#define GMPI_RECV_BUF_LEN 65536
-static int _gmpi_parse_init_recv_msg(srun_job_t *job, char *rbuf,
- gm_slave_t *slave_data);
-
static int gmpi_fd = -1;
static int _gmpi_parse_init_recv_msg(srun_job_t *job, char *rbuf,
- gm_slave_t *slave_data)
+ gm_slave_t *slave_data, int *ii)
{
unsigned int magic, id, port_board_id, unique_high_id,
unique_low_id, numanode, remote_pid, remote_port;
@@ -76,6 +73,7 @@ static int _gmpi_parse_init_recv_msg(srun_job_t *job, char *rbuf,
got = sscanf(rbuf, "<<<%u:%u:%u:%u:%u:%u:%u::%u>>>",
&magic, &id, &port_board_id, &unique_high_id,
&unique_low_id, &numanode, &remote_pid, &remote_port);
+ *ii = id;
if (got != 8) {
error("GMPI master received invalid init message");
return -1;
@@ -113,8 +111,9 @@ static int _gmpi_parse_init_recv_msg(srun_job_t *job, char *rbuf,
static int _gmpi_establish_map(srun_job_t *job)
{
struct sockaddr_in addr;
+ in_addr_t *iaddrs;
socklen_t addrlen;
- int accfd, newfd, rlen, nprocs, i, j;
+ int accfd, newfd, rlen, nprocs, i, j, id;
size_t gmaplen, lmaplen, maplen;
char *p, *rbuf = NULL, *gmap = NULL, *lmap = NULL, *map = NULL;
char tmp[128];
@@ -127,6 +126,7 @@ static int _gmpi_establish_map(srun_job_t *job)
accfd = gmpi_fd;
addrlen = sizeof(addr);
nprocs = opt.nprocs;
+ iaddrs = (in_addr_t *)xmalloc(sizeof(*iaddrs)*nprocs);
slave_data = (gm_slave_t *)xmalloc(sizeof(*slave_data)*nprocs);
for (i=0; i<nprocs; i++)
slave_data[i].defined = 0;
@@ -147,8 +147,11 @@ static int _gmpi_establish_map(srun_job_t *job)
} else {
rbuf[rlen] = 0;
}
- if (_gmpi_parse_init_recv_msg(job, rbuf, slave_data) == 0)
+ if (_gmpi_parse_init_recv_msg(job, rbuf, slave_data,
+ &id) == 0) {
i++;
+ iaddrs[id] = ntohl(addr.sin_addr.s_addr);
+ }
close(newfd);
}
xfree(rbuf);
@@ -184,9 +187,7 @@ static int _gmpi_establish_map(srun_job_t *job)
dp = &slave_data[i];
p = lmap;
for (j=0; j<nprocs; j++) {
- int jhostid = step_layout_host_id (job->step_layout, j);
-
- if ((ihostid == jhostid) &&
+ if (iaddrs[i] == iaddrs[j] &&
(dp->numanode == slave_data[j].numanode)) {
sprintf(tmp, "<%u>", j);
strcpy(p, tmp);
@@ -212,8 +213,7 @@ static int _gmpi_establish_map(srun_job_t *job)
error("setsockopt in GMPI master: %m");
bzero(&addr, sizeof(addr));
addr.sin_family = AF_INET;
- addr.sin_addr.s_addr
- = job->slurmd_addr[ihostid].sin_addr.s_addr;
+ addr.sin_addr.s_addr = htonl(iaddrs[i]);
addr.sin_port = htons(dp->remote_port);
if (connect(newfd, (struct sockaddr *)&addr, sizeof(addr)))
fatal("GMPI master failed to connect");
@@ -224,6 +224,7 @@ static int _gmpi_establish_map(srun_job_t *job)
xfree(slave_data);
xfree(lmap);
xfree(gmap);
+ xfree(iaddrs);
debug2("GMPI master responded to all GMPI processes");
return 0;
diff --git a/src/srun/io.c b/src/srun/io.c
index c95658a7147..ffdd121405e 100644
--- a/src/srun/io.c
+++ b/src/srun/io.c
@@ -728,9 +728,9 @@ again:
list_enqueue(server->msg_queue, msg);
}
} else if (header.type == SLURM_IO_STDIN) {
- debug("SLURM_IO_STDIN");
int nodeid;
struct server_io_info *server;
+ debug("SLURM_IO_STDIN");
msg->ref_count = 1;
nodeid = step_layout_host_id(info->job->step_layout,
header.gtaskid);
--
GitLab