> On April 17, 2017, 3:09 p.m., Jay Guo wrote: > > src/master/allocator/sorter/drf/sorter.cpp > > Line 333 (original), 509 (patched) > > <https://reviews.apache.org/r/57254/diff/17/?file=1692350#file1692350line536> > > > > In case of `dirty`, how about insert logic of `listClients` there to > > avoid traversing tree twice for the sake of performance? I tried a > > benchmark test which does suggest a performance benefit.
That could be a useful improvement. When you benchmarked this, what % performance improvement did you observe? - Neil ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/57254/#review172083 ----------------------------------------------------------- On April 13, 2017, 9:42 p.m., Neil Conway wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/57254/ > ----------------------------------------------------------- > > (Updated April 13, 2017, 9:42 p.m.) > > > Review request for mesos, Benjamin Bannier, Benjamin Mahler, and Michael Park. > > > Repository: mesos > > > Description > ------- > > This commit replaces the sorter's flat list of clients with a tree; the > tree represents the hierarchical relationship between sorter clients. > Each node in the tree contains a vector of pointers to child nodes. The > tree might contain nodes that do not correspond directly to sorter > clients. For example, adding clients "a/b" and "c/d" results in the > following tree: > > root > -> a > -> b > -> c > -> d > > The `sort` member function still only returns one result for each active > client in the sorter. This is implemented by ensuring that each sorter > client is associated with a leaf node in the tree. Note that it is > possible for a leaf node to be transformed into an internal node by a > subsequent insertion; to handle this case, we "implicitly" create an > extra child node, which maintains the invariant that each client has a > leaf node. For example, if the client "a/b/x" is added to the tree > above, the result is: > > root > -> a > -> b > -> . > -> x > -> c > -> d > > The "." leaf node holds the allocation that has been made to the "a/b" > client itself; the "a/b" node holds the sum of all the allocations that > have been made to the subtree rooted at "a/b", which also includes > "a/b/x". > > This commit also introduces a new approach to sorting: rather than > keeping an `std::set` of sorter clients, we now keep a tree of > `std::vector`, which is sorted explicitly via `std::sort`. The previous > implementation tried to optimize the sorting process by updating the > sort order incrementally when a single sorter client was updated; this > commit removes that optimization, and instead re-sorts the entire tree > whenever the sort order is changed. > > Re-introducing a version of this optimization would make sense in the > future (MESOS-7390), but benchmarking suggests that this commit results > in a net improvement in sorter performance anyway. The sorter perf > improvement is likely due to the introduction of a secondary hashmap > that allows the leaf node associated with a tree path to be find > efficiently; the previous implementation required a linear scan of a > `std::set`. > > > Diffs > ----- > > src/master/allocator/sorter/drf/metrics.cpp > 15aab32db5ca1a7a14080e9bbb7c65283be3ec20 > src/master/allocator/sorter/drf/sorter.hpp > 76329220e1115c1de7810fb69b943c78c078be59 > src/master/allocator/sorter/drf/sorter.cpp > ed54680cecb637931fc344fbcf8fd3b14cc24295 > src/master/allocator/sorter/sorter.hpp > b3029fcf7342406955760da53f1ae736769f308c > src/tests/hierarchical_allocator_tests.cpp > 33e7b455f8664858eb4f03727b076a10c80cd6e0 > src/tests/master_allocator_tests.cpp > 119e318f8a01d50e8dae5c30cf5fa6a017c3c625 > src/tests/sorter_tests.cpp 43bd85798aef0c89751b725ebf35308a5e9e997a > > > Diff: https://reviews.apache.org/r/57254/diff/17/ > > > Testing > ------- > > > Thanks, > > Neil Conway > >
