MPIEXEC 0.75 + PBSPro ?

[Pete Wyckoff]

francois at hpce.nec.com said on Thu, 25 Mar 2004 21:24 +0100:
> i just grab the last mpiexec version (0.75) and i was trying to have it
> working with PBS Pro(5.4).
> 
> As reported in the mpiexec web page, notes section, PBS Pro seems  not to
> work correctly with mpiexec.
> 
> I'm getting the following error messages  when trying to run a 2 cpus job
> (-l nodes=2:ppn=1)
> 
> mpiexec: Warning: get_hosts: ncpus=2 but nodect=2, pretending nodect=1.
> mpiexec: Warning: get_hosts: ncpus=2 but numtask=2, expecting numtask=1.
> 
> The job ran only on 1 cpu.
> 
> Did somebody already encounter this issue and fix it ?

Yes, sorry, fixed on February 4, but after the 0.75 release.  The CVS
version should be fine.  I updated the web page and will roll another
release soon.  Here's a diff for get_hosts.c you could add by hand.

Let us know if there are any other issues.

		-- Pete
-------------- next part --------------
Index: get_hosts.c
===================================================================
RCS file: /cvs/mpiexec/get_hosts.c,v
retrieving revision 1.36
retrieving revision 1.37
diff -u -r1.36 -r1.37
--- get_hosts.c	8 Dec 2003 23:02:34 -0000	1.36
+++ get_hosts.c	4 Feb 2004 14:22:40 -0000	1.37
@@ -1,6 +1,6 @@
 /*
  * get_hosts.c - read hostnames from pbs, mark which ones we'll use
- * $Id: get_hosts.c,v 1.36 2003/12/08 23:02:34 pw Exp $
+ * $Id: get_hosts.c,v 1.37 2004/02/04 14:22:40 pw Exp $
  *
  * Copyright (C) 2000-3 Ohio Supercomputer Center.
  * Distributed under the GNU Public License Version 2 or later (See LICENSE)
@@ -224,30 +224,40 @@
     /* close connection to pbs server */
     pbs_disconnect(fd);
 
+    /*
+     * Various PBSes disagree about what should appear here.  Try to do
+     * the best thing.
+     *   OpenPBS, non-ts: nodes=20:ppn=2 nodect=20
+     *   PBSPro, not-ts:  nodes=20:ppn=2 nodect=20 ncpus=40
+     *   OpenPBS, ts:     ncpus=2
+     *   PBSPro, ts:      ??
+     */
     if (!(have_ncpus || have_nodect))
 	error("%s: pbs_statjob returned neither \"ncpus\" nor \"nodect\"",
 	  __func__);
     if (have_ncpus > 1) {
-	task_cntrl_t *oldtasks = tasks;
-
-	if (have_nodect > 1)
-	    warning("%s: ncpus=%d but nodect=%d, pretending nodect=1",
-	      __func__, have_ncpus, have_nodect);
-	if (numtask > 1)
-	    warning("%s: ncpus=%d but numtask=%d, expecting numtask=1",
-	      __func__, have_ncpus, numtask);
-	/*
-	 * Explode multi-cpu task entries into 1-cpu ones for better
-	 * config matching.  Later these will be compressed back down
-	 * for spawning.
-	 */
-	numtask = have_ncpus;
-	tasks = Malloc(numtask * sizeof(*tasks));
-	for (i=0; i<numtask; i++) {
-	    memcpy(&tasks[i], &oldtasks[0], sizeof(*tasks));
-	    tasks[i].name = strsave(oldtasks[0].name);
+	if (cl_args->verbose > 2) {
+	    printf("%s: numtask=%d ncpus=%d nodect=%d\n", __func__, numtask,
+	      have_ncpus, have_nodect);
+	}
+	if (have_nodect > 1 || numtask > 1) {
+	    /* ignore the ncpus setting, trust nodect and exec_host */
+	    ;
+	} else {
+	    /*
+	     * Explode multi-cpu task entries into 1-cpu ones for better
+	     * config matching.  Later these will be compressed back down
+	     * for spawning.
+	     */
+	    task_cntrl_t *oldtasks = tasks;
+	    numtask = have_ncpus;
+	    tasks = Malloc(numtask * sizeof(*tasks));
+	    for (i=0; i<numtask; i++) {
+		memcpy(&tasks[i], &oldtasks[0], sizeof(*tasks));
+		tasks[i].name = strsave(oldtasks[0].name);
+	    }
+	    free(oldtasks);
 	}
-	free(oldtasks);
     }
 
     /* enforce one process per physical node by strcmp on host name */
@@ -260,9 +270,10 @@
     if (cl_args->numproc) {
 	if (cl_args->numproc > numtask)
 	    error(
-	     "%s: argument -n specifies %d processors, only %d available%s",
+	     "%s: argument -n specifies %d processors, but\n"
+	     "  only %d available%s",
 	      __func__, cl_args->numproc, numtask,
-	      cl_args->pernode ? "\n  after processing -pernode flag" : "");
+	      cl_args->pernode ? " after processing -pernode flag" : "");
 	/* just take whatever the user specified, which may be fewer,
 	 * discarding the rest of the tasks */
 	numtask = cl_args->numproc;