Re: [BProc] bjs not sane

[Luke P. <lo...@du...>]

Here's a diff for shared.c that I think does the trick.  Just an
uninitialized variable and a misplaced control structure.

79c79
<     int i, node_count;
---
>     int i, node_count=0;
95,97d94
<       job_counts[node_count].node  = node->node;
<       job_counts[node_count].count = 0;
<
99a97,98
>           job_counts[node_count].node  = node->node;
>           job_counts[node_count].count = 0;



-Luke

On Tue, 2004-06-08 at 14:48, er...@he... wrote:
> On Mon, Jun 07, 2004 at 05:59:04AM -0500, Luke Palmer wrote:
> > Hi everyone,
> > 
> > I have a bproc cluster up and running of Fedora core 1.  bpsh looks to
> > be working.  bjs doesn't appear to be doing anything sane, however.  I'm
> > using v1.5.
> > 
> > In my test configuration I have 4 SMP nodes.  Using the "filler" policy,
> > bjs will place 4 processes on the master only.  Slaves do not get used.
> 
> This is the way that bjs is supposed to operate.  On a BProc system
> placing processes on the slave nodes is normally handled by something
> like mpirun.  Jobs usually look like simple scripts (which run on the
> front end) that contain commands to start stuff on the slave nodes.
> The NODES environement variable should tell the script where to run
> things.
> 
> > "simple" will start as many jobs as I want, but only on the master. 
> > bjsctl -r does not kill a running job, either.
> 
> bjsctl -r will kill off stuff running on the nodes allocated to the job.
> 
> > "shared" just plain doesn't work.  Any call to bjssub gives me this:
> > 
> > connect("/tmp/.bjs"): Connection refused
> > Failed to connect to scheduler.
> > Is bjs not running or is /tmp/.bjs the wrong socket path?
> 
> That usually means bjs is died for some reason.  It could be that
> "shared" is busted.  That one doesn't get a lot of testing around
> here.
> 
> - Erik