[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Condor-users] dagman progress as classad?

Hi Kent,
Horvath, are you coming to Condor Week?  Himani Apte is going to talk
about her research in incorporating ClassAds into DAGMan.
I won't be at the Condor Week but I'd be most interested in this research.
We're using dagman on windows extensively (and exclusively) and I'm very curious about possible additions.

What she has done so far doesn't exactly address what you're interested
in, but it would certainly make it easier to do what you want.  Himani is
interested in hearing what people want to do with DAGMan to take advantage
of the capabilities of ClassAds.
The three most important things I miss from dagman are:

- Inter dag-job communication. At the moment there is no way to communicate between dependent dag jobs, the only solution is modifying the child's submit file after a parent job completes. A problem with this, is that if a parent job is restarted after completion there is no way to restart the child jobs without serious condor_qedit hacking, since the submit file modifications are not propagated into the job queue. If we had classAd attributes on the parent node that could be updated on a job and child jobs could read them (possibly through environment variables) than these job connections worked
much simpler and no file modification was required.

- Dagman status attributes, something along the lines of my previous mail.

- L:icense management at the dagman level: don't submit jobs if there is no sufficient license since jobs would just sit in the queue and force pointless negotiation. Needs a license management framework though.