> On 六月 21, 2016, 1:23 a.m., Benjamin Mahler wrote: > > include/mesos/mesos.proto, lines 278-281 > > <https://reviews.apache.org/r/48914/diff/1/?file=1423477#file1423477line278> > > > > How about the following? > > > > ``` > > // Indicates whether the framework is aware of GPU resources. > > // Frameworks that are aware of GPU resources are expected to > > // avoid placing non-GPU workloads on GPU agents, in order > > // to avoid occupying a GPU agent and preventing GPU workloads > > // from using precious GPU resources! Currently, if a > > // framework is unaware of GPU resources, it will not be > > // offered *any* of the resources on an agent with GPUs. > > // This restriction is in place because we do not have a > > // revocation mechanism that ensures GPU workloads can evict > > // GPU agent occupants if necessary. > > // > > // TODO(bmahler): As we add revocation we can relax the > > // restriction here. See MESOS-5634 for more information. > > ```
@Ben, I think that here should be "// to avoid occupying a GPU agent and preventing `non-GPU` workloads"? - Guangya ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/48914/#review138698 ----------------------------------------------------------- On 六月 18, 2016, 10:07 p.m., Kevin Klues wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/48914/ > ----------------------------------------------------------- > > (Updated 六月 18, 2016, 10:07 p.m.) > > > Review request for mesos and Benjamin Mahler. > > > Bugs: MESOS-5634 > https://issues.apache.org/jira/browse/MESOS-5634 > > > Repository: mesos > > > Description > ------- > > Due to the scarce resource problem described in MESOS-5377, we are > introducing a GPU_RESOURCES Framework capability. This capability > allows the Mesos allocator to make better decisions about which > frameworks should receive resources from GPU capable machines. In > essence, the allocator ONLY allocate resources from GPU capable > machines to frameworks that have this capability. This is necessary to > prevent non-GPU workloads from filling up the GPU machines and > preventing GPU workloads to run. > > > Diffs > ----- > > include/mesos/mesos.proto e4c5bd31cf035707036eb509336fe051119b4e78 > include/mesos/v1/mesos.proto 9be22f02861f1eb89ab547d88530faf90ebee7ab > src/master/allocator/mesos/hierarchical.hpp > 9c6b23abe2b0cb16412f1ed90165f8d0c14552fa > src/master/allocator/mesos/hierarchical.cpp > 8b7b3afb5770c617918ec4864faaf8d8a7a81ef2 > src/tests/containerizer/nvidia_gpu_isolator_tests.cpp > e06d107f2dcdb9b470e330c8ceee66a54220d41b > > Diff: https://reviews.apache.org/r/48914/diff/ > > > Testing > ------- > > $ make -j check; sudo GTEST_FILTER="*NVIDIA*" src/mesos-tests > > > Thanks, > > Kevin Klues > >
