[torqueusers] pbsdsh implementation bug?

Garrick Staples garrick at usc.edu
Fri Jan 4 16:35:15 MST 2008


On Fri, Jan 04, 2008 at 04:45:19PM +0000, Marcin Mogielnicki alleged:
> Hi all,
> 
> The case is that aux/JOBID entries does not work when used with 'pbsdsh -h'.
> 
> I'm trying to use pbsdsh -h option and I discovered that pbdsh names 
> from aux/JOBID are not used. What pbsdsh does is grepping the name out 
> of 'uname -a' output for every node and comparing it to -h argument. So 
> implication is that hostname must be exactly the same as label given in 
> torque nodes file while it is often not true. Well, if I want to execute 
> something on the second node assigned for example I definitely expect 
> the second entry from aux/JOBID file to work...
> 
> It doesn't look like big problem, but in fact it sometimes is. I 
> encountered bunch of commercial applications requiring defining command 
> for executing anything on remote nodes. For example 'rsh -l %U %H' 
> should be substituted by 'pbsdsh -h %H' in one specific case. 
> Application takes node names from pbs. Crash is guaranteed here.
> 
> It looks like programmer's shortcut for me, as proper long operations on 
> bunch of structures, needed to catch pbs defined node name, were 
> replaced by few lines only based on mostly right assumption. Mostly - 
> but not always. What I'm interested in is confirming if it is considered 
> buggy (i.e. design flaw) behaviour at all. Are there any plans to fix it?

Yes, it was a quick solution.  Looking at the uname output was the only way
that pbsdsh could see the hostname since it is never actually passed the
hostlist.

The real solution is entirely non-trivial requiring TM protocol changes.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://www.supercluster.org/pipermail/torqueusers/attachments/20080104/b4347f6b/attachment.bin


More information about the torqueusers mailing list