On 08/11/2015 02:21 PM, Jens Schmaler wrote: > In my case, the executable is the same for all jobs in one cluster, but > it is different for each cluster. So a separate class ad does not really > seem to be a viable solution, right? Ah, I see -- I missed the "different for each cluster" bit. One problem with external file transfer mechanisms is the race condition between when the transfer finishes and when the job starts. So having the transfer as parent node in the dag is a way around that. But then you have to lay it out so the transfer is once per machine and in your case also once per cluster. You can probably make cluster n+1 transfer and jobs not start until cluster n transfer's finished, too, by making it a child of transfer n. -- Dimitri Maziuk Programmer/sysadmin BioMagResBank, UW-Madison -- http://www.bmrb.wisc.edu
Attachment:
signature.asc
Description: OpenPGP digital signature