Hey Raymond,

Glad to hear that you are interested in Mesos. Please see my answers inline.


It specifically is talking about resource requirements at the framework
> level.
> What if some tasks in the one framework require a GPU and others do not ?
> The kind of resources that tasks from Beaker require are not at all
> homogeneous.
> So while one task may require a GPU, the other may require instead a
> specific SCSI card.
>
>
Non-homogenous task requirements for a framework are definitely ok/expected
in the Mesos world.
The resources that each slave offers can be set by the operator as command
line flags (--resources) on the slaves.
It is up to the framework to chose among the resources offered according to
its tasks' requirements.


> Can you have some tasks be rigid though and some other not? Or is it
> strictly at the framework
> level. I'm thinking of the case where task A, B and C are rigid and
> co-dependent
> on each other, D and E are rigid and dependent on each-other, but task F
> is elastic.
> If this is not possible, perhaps Beaker could also present A, B, and C as
> a single task, but require
> three 'slots'. Maybe.
>
>
Not sure what your "rigid and co-dependent" requirements are, but if they
have to be running on the same slave and need a strict ordering it would be
better to wrap them in a single task and explicitly control their
lifecycle. If the rigidity is more in the sense of all 3 tasks need to be
run (on different hosts) or not, the framework can hold on to enough
resource offer slots before scheduling all 3 tasks. The caveat here is that
a framework holding on to offers doesn't play nicely with DRF and might
starve other frameworks in the cluster.

Reply via email to