Re: [BProc] mpirun

[Nicholas H. <he...@se...>]

On Fri, 18 Oct 2002 12:47:33 -0400
er...@he... wrote:
 -- is this normal, or is it 'fixable'?
> 
> It's normal and fixable.  The issue is that the cruddy I/O forwarding
> hack only works for processes created directly from the front end.  In
> the tree spawn with vrfork() only rank 0 is created directly from the
> front end.

Gotcha -- I am working on that now -- using bpsh's base functions.

> 
> The fix would be to add a bpsh-like I/O forwarder to mpirun and do a
> similar I/O thing.
> 
> - Erik
Here is a patch to catch an error of /etc/beowulf/nodeinfo is missing:

bash-2.05$ diff -urN /home/henken/src/bproc/mpirun-0.2/mpirun.c mpirun.c
--- /home/henken/src/bproc/mpirun-0.2/mpirun.c  Tue Oct 22 15:33:06 2002
+++ mpirun.c    Tue Oct 22 15:30:26 2002
@@ -101,7 +101,11 @@

     list_size = bproc_numnodes();

-    f = fopen(dbfile, "r");
+    if ((f = fopen(dbfile, "r")) == NULL) {
+               fprintf(stderr,"could not open %s\n", dbfile);
+               perror("fopen");
+               exit(1);
+       }
     lk.l_type   = F_RDLCK;
     lk.l_whence = SEEK_SET;
     lk.l_start  = 0;

And: shouldn't bproc_vexecmove_io fail of prog points a nonexistant
program?

cheers
Nic
-- 
Nicholas Henke
Linux Cluster Systems Programmer
he...@se... - 215.573.8149