[Bigjob-users] BigJob Condor / [BigJob] Unable to control SPMDVariation during pilot_job submission (#1)

[Andre Luckow]

Hi Ole,
I implemented a workaround for the spmd_variation in BJ 0.4.18 => it
is not set for Condor URLs.

>> 1) We use the URL scheme condorg://<hostname> to trigger the Condor
>> specific BJ plugin.
>
> that's the condor-G case. However, OSG prefers "vanilla" condor (condor://localhost) job submission, since this is backed by their glidein-WMS system. Using Condor-G directly is discouraged since it overrides their fair-share policies and confuses the schedulers.

Understand. I am not an expert in the Glide-In WMS stuff. Is GlideIn
WMS offered as a service?
Looking at the documentation, one could think of writing an adaptor
for Glide-WMS:

start_pilot_job() => Submits a condor startd wrapper to grid sites via
Glide-In Factory.
subjob submission => Submit job to Condor pool

But, that would mean we need a Glide-In WMS SAGA adaptor and not a
Condor adaptor.

>> 2) We have a special version of the agent
>> bigjob/bigjob_agent_condor.py (not sure whether we need this - since
>> most of the code is the same).
>
> Hm. So what is it for then? What was it used for and on what platform?
I think the version we have in the SVN is not the most current one
(hopefully Sharath can help).

>> 3.) When bigjob.start_pilot_job is called we submit a Condorg job via
>> SAGA/Condor which starts the BJ agent
>
> Again, we could. But: Condor-G is not the way to do it on OSG.
>
>> 4.) Sub-Jobs/Work-Units are spawned via the BJ agent - there is no
>> Condor pool that is actually used.
>
> In the Condor-G case I assume the answer is 'yes'. In the 'vanilla' case: I have *no* idea how that is supposed to work.

So, I guess we have to decide what way we want to go:

1.) Condor-G seems to me the more feasible option, but we wouldn't use
the real Glide-In in this case. We would just use Condor to spawn BJ.
2.) GlideIn-WMS probably would mean to wrap the WMS commands in a
BJ/Troy-style APIs. Since, we wouldn't use the BJ agent in this case,
it probably doesn't make to much sense to include as a plugin into BJ
(although we certainly could do that).

Best,
Andre