[Swift-devel] Problems with coasters and managedfork jobmanager

[Michael Wilde]

On 2/6/09 8:04 AM, Ben Clifford wrote:
> On Thu, 5 Feb 2009, Michael Wilde wrote:
> 
>> I see Mats's note in a prio mail about concern that the managed-fork mechanism
>> may kill the coaster service, but no comments about script parsing errors.
> 
> The condor jobmanager deals quite poorly with whitespace in arguments, in 
> a way that I cannot see how to work around. (I've run into a very similar 
> problem when looking at making Swift run without any shared filesystem).
> 
> This bit almost definitely doesn't work with existing jobmanager-condor.
> 
>>            js.setExecutable("/bin/bash");
>>            js.addArgument("-c");
>>            js.addArgument(loadBootstrapScript());
> 

I see. The problem turns out to be the newlines in the command script. 
It can be reproduced with globusrun:

com$ globusrun -o -r 
tp-grid1.ci.uchicago.edu:2119/jobmanager-managedfork 
'&(executable="/bin/echo") (arguments= "hello world")'
hello world


com$ globusrun -o -r 
tp-grid1.ci.uchicago.edu:2119/jobmanager-managedfork 
'&(executable="/bin/echo") (arguments= "hello
 >  world")'

ERROR: Failed to parse command file (line 10).
GRAM Job failed because the job failed when the job manager attempted to 
run it (error code 17)
com$

--

I'll make a brief attempt to work around this, but most likely wont be 
able to, as you say.

- Mike


> GRAM provided an update package to VDT/OSG the other day that changes 
> condor jobmanager whitespace handling so that it may be possible to make 
> it work. See this thread: 
> http://lists.globus.org/pipermail/gram-user/2009-January/000790.html
> 
> With the present deployed infrastructure, one approach might be to have 
> the bootstrap script staged in as a file using file transfer mechanisms 
> (in the quickest hack case, staged in at the same time as wrapper.sh and 
> seq.sh by swift, though this will not work if you are trying to use the 
> coaster filesystem provider), allowing the shell command to have spaces 
> removed.
>