> On Aug. 7, 2014, 3:39 p.m., Adam B wrote: > > Have you done any simulations or performance tests on a test cluster? I'd > > love to see some numbers that prove that this is more fair in some scenario > > than the previous algorithm. Not sure I can blindly trust a unit test.
Not yet. I'll have to think about how to test "fairness". Intuitively though, this is doing more fine-grained resource allocation. Previously we were giving roughly the entire cluster's resources to one framework with the smallest share and waiting for it decline them in order to give them to another framework. With this change, a framework will be allocated enough slaves' resources to make its DRF share on par/better than other frameworks. This way multiple frameworks would be given "some" resources during every allocation, while still honoring DRF. It would also help ameliorate concerns around frameworks holding on to offers and not playing well with DRF. - Vinod ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/24356/#review49901 ----------------------------------------------------------- On Aug. 7, 2014, 11:02 p.m., Vinod Kone wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/24356/ > ----------------------------------------------------------- > > (Updated Aug. 7, 2014, 11:02 p.m.) > > > Review request for mesos, Adam B and Benjamin Hindman. > > > Bugs: MESOS-1119 > https://issues.apache.org/jira/browse/MESOS-1119 > > > Repository: mesos-git > > > Description > ------- > > The allocator now updates the DRF algorithm after every allocation of a > slave's resources. This will make the allocation more granular and fair. > > > Diffs > ----- > > src/master/hierarchical_allocator_process.hpp > d81082f7fa2a4ffc13f18c232abb8ad2814bebc1 > src/tests/allocator_tests.cpp f0226cb3fa4b54f917c0a0f59b986e9115352507 > > Diff: https://reviews.apache.org/r/24356/diff/ > > > Testing > ------- > > make check > > make check GTEST_FILTER="*PerSlaveAllocation*" GTEST_REPEAT=1000 > GTEST_BREAK_ON_FAILURE=1 GTEST_VERBOSE=1 MESOS_VERBOSE=1 GLOG_v=1 > > > Thanks, > > Vinod Kone > >
