[ https://issues.apache.org/jira/browse/MESOS-8989?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16511798#comment-16511798 ]
Meng Zhu commented on MESOS-8989: --------------------------------- So currently, range subtraction is about 2-3 times more expensive than addition. > Add a better benchmark for range type resources. > ------------------------------------------------ > > Key: MESOS-8989 > URL: https://issues.apache.org/jira/browse/MESOS-8989 > Project: Mesos > Issue Type: Improvement > Reporter: Meng Zhu > Assignee: Meng Zhu > Priority: Minor > Labels: benchmark > > While investigating the performance issue with the `available()` function in > the allocator, the arithmetic of range type resources, namely ports, appear > to be most expensive. And the subtraction proves to be significantly more > expensive than the addition whose cause is not immediately clear to me. > While we do have a benchmark > `[Resources_Scalar_Arithmetic_BENCHMARK_Test|https://github.com/apache/mesos/blob/1b851877ff22cf2da193bb649dc838bb50c26b34/src/tests/resources_tests.cpp#L3610-L3626]` > that contains a range parametrized test, its result is somewhat misleading. > It evaluates the addition performance by repeatedly summing up the same port > resources which is fine. However, when evaluating the subtraction, it took > the result from the addition and repeatedly subtracting the same port > resources. The result is that after the first subtraction, the subtrahend > becomes empty and all the subsequent subtractions are shorthanded, producing > uninstructive result. -- This message was sent by Atlassian JIRA (v7.6.3#76005)