> > There has been separate discussion around supporting arbitrary resources. > Is it plausible to build that and get GPU support for free?
I am certainly open to generalizing internal resource management as much as possible as I envision us adding more resource types in the future (e.g. network as a bandwidth). That said, I was intending to stick with explicit fields at the TaskConfig level. Are you suggesting going as far as generalizing resource representation at the thrift level similar to how Mesos handles it: https://github.com/apache/mesos/blob/master/include/mesos/mesos.proto#L631? That would require a much bigger effort and longer deprecation cycle but I'd be happy to explore it as well. On Wed, Apr 6, 2016 at 10:59 AM, Joshua Cohen <jco...@apache.org> wrote: > +1 > > On Wed, Apr 6, 2016 at 12:41 PM, Maxim Khutornenko <ma...@apache.org> > wrote: > > > Mesos community is finalizing their MVP for supporting GPUs: > > https://issues.apache.org/jira/browse/MESOS-4424 > > Design doc: > > > https://docs.google.com/document/d/10GJ1A80x4nIEo8kfdeo9B11PIbS1xJrrB4Z373Ifkpo/edit > > > > Would anyone have any reservations about supporting GPUs in Aurora? > > Initial walk through our resource management codebase shows we can do > > it without too many changes or excessive refactoring. If there are no > > objections I am willing to put up a design doc for review next. > > > > Thanks, > > Maxim > > >