Flink doesn't support a 'unique host' constraint at this time; it simply
accepts adequate offers without any such consideration. Flink does
support a 'host attributes' constraint to filter certain hosts, but that's
not applicable here.
Under the hood, Flink uses a library called Netflix Fenzo to optimize
placement, and a uniqueness constraint could be added by more deeply
leveraging Fenzo's constraint system. You mentioned that you're trying to
make good use of your GPU resources, which could also be achieved by
treating GPU as a scalar resource (similar to how memory and cores are
treated). Mesos does support that, but Fenzo may require some
enhancement. So, these are two potential ways to enhance Flink to support
your scenario. I'm happy to help; reach out to me.
The obvious, ugly workaround is to configure your TMs to be large enough to
consume the whole host.
On Thu, Jan 11, 2018 at 7:18 AM, Gary Yao <[EMAIL PROTECTED]> wrote: