[Swift-devel] provider staging question
Jonathan Monette
jonmon at mcs.anl.gov
Tue Aug 16 21:14:24 CDT 2011
That is what I was thinking about CDM but wasn't sure if that is what you meant in the email a couple days ago.
I thought that the scratch tag was broken from the sites file. I thought it was mentioned before that Papia was experiencing problems when using the scratch tag in her sites file.
On Aug 16, 2011, at 8:15 PM, Michael Wilde wrote:
> Hi Jon,
>
> I'll try a short answer, but this needs more thoight, and much testing:
>
> - Dont use CDM with provider staging (yet; maybe someday that will make sense...)
> I dont tthink the two will work together well.
>
> - ps should stage to the local hard disk; I *think* it may honor the workdirectory tag as where to stage, though. Or maybe the scratch tag? I think it defaults to /tmp. This needs to be tested and documented. Mihael, can you clarify?
>
>
> Miek
>
>
> ----- Original Message -----
>> From: "Jonathan Monette" <jonmon at mcs.anl.gov>
>> To: "Michael Wilde" <wilde at mcs.anl.gov>
>> Sent: Tuesday, August 16, 2011 6:08:47 PM
>> Subject: provider staging question
>> Mike,
>> I am configuring the run to test the SwiftMontage runs with provider
>> staging turned on as you suggested. When I turn on provider staging
>> should I not use CDM? The issue that I was experiencing that led me to
>> start using CDM was extremely long copy times from the cwd to the job
>> directory specified by the <workdirectory> in the sites.xml file. Does
>> provider staging circumvent that issue? I know provider staging copies
>> the input files directly onto the compute nodes local disk but that is
>> about all I know. Could you fill me in a bit on what exactly provider
>> staging does? Would using both CDM "direct" directives and provider
>> staging cause some degraded performance?
>
> --
> Michael Wilde
> Computation Institute, University of Chicago
> Mathematics and Computer Science Division
> Argonne National Laboratory
>
More information about the Swift-devel
mailing list