> On June 14, 2016, 10:50 p.m., Jiang Yan Xu wrote: > > src/master/allocator/sorter/drf/sorter.cpp, lines 457-467 > > <https://reviews.apache.org/r/45961/diff/3/?file=1390649#file1390649line457> > > > > For this we need to > > > > 1. Run existing benchmarks to see how much performance degradation is > > there for clusters (large number of agents or frameworks) that are not > > using shared resources. Ideally there should be negligible performance > > impact for such clusters. > > 2. Create a benchmark that uses shared resources in a realistic way. > > > > In implementing this patch we do have the options to store shared > > resources separately in a few places (e.g., store shared resources outside > > of `allocations[name].resources`) that can help with performance but the > > trade off is that it may increase code complexity. Benchmarks can help us > > address these tradeoffs.
Verified that changes in allocator and sorter does not affect run time performance adversely for regular resources. I will add a benchmark for shared resources in the next update. - Anindya ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/45961/#review136970 ----------------------------------------------------------- On June 23, 2016, 3:50 p.m., Anindya Sinha wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/45961/ > ----------------------------------------------------------- > > (Updated June 23, 2016, 3:50 p.m.) > > > Review request for mesos, Benjamin Mahler, Joris Van Remoortere, Michael > Park, and Jiang Yan Xu. > > > Bugs: MESOS-4431 > https://issues.apache.org/jira/browse/MESOS-4431 > > > Repository: mesos > > > Description > ------- > > Each shared resource is accouted via its share count. This count is > updated based on the resource operations (such as add and subtract) > in various scenarios such as task launch and terminate at multiple > modules such as master, allocator, sorter, etc. > Only allow DESTROY if no task is using the volume, or has not been > offered to any framework. > Since shared resources are available to multiple frameworks of the > same role simultaneously, we normalize it with a weight equivalent > to the number of frameworks to which this shared resource has been > allocated to in the sorter. > > > Diffs > ----- > > src/common/resources.cpp f6ff92b591c15bc8e93fd85e1896349c3a7bb968 > src/common/resources_utils.cpp 8e881a09500b0966b577e514077460f724868a8d > src/master/allocator/mesos/hierarchical.cpp > 5b2331b3f99252fd16f2514123006dd4ba7f1c0d > src/master/allocator/sorter/drf/sorter.hpp > 35273b5dcf39938125a112c5e56b5a8a542157db > src/master/allocator/sorter/drf/sorter.cpp > 27d56f274c41bbabe6f2abbbcebd2c4eff52693e > src/master/http.cpp 7daaf12a4086635bbc5aba5e3375c95e8899ac6e > src/master/master.hpp fe57878dc59637459d5c5cdae0be2aa159133fa4 > src/master/master.cpp 8def7156f4a05b39590321ce7743f7385a68bed0 > src/master/validation.hpp 43d876b00511d4ef652c2685d4950c994b5eef31 > src/master/validation.cpp 9120b71fc7725bdf7094aac6619d8aadcc352df5 > src/tests/master_validation_tests.cpp > 9eb82a569f7c48caa96d4d54a93b199ccac74385 > src/tests/sorter_tests.cpp 6fc58829892dc0223140f1b47593a3e5853cace5 > src/v1/resources.cpp 8c3f2d1c1529915a59d47fe37bb3fc7a3267079a > > Diff: https://reviews.apache.org/r/45961/diff/ > > > Testing > ------- > > Tests successful. > > > Thanks, > > Anindya Sinha > >
