> On June 13, 2016, 12:02 p.m., Joris Van Remoortere wrote: > > src/common/values.cpp, lines 409-420 > > <https://reviews.apache.org/r/48593/diff/2/?file=1415977#file1415977line409> > > > > Thanks for digging into this issue further. > > Please do read my response on your e-mail thread for further context. > > > > This patch looks like it fixes a particular bottleneck by converting to > > intervalset inline. > > I understand why you did this, as my e-mail outlines that changing the > > stored data-type is more complicated. > > > > I can see how this improves performance; however, is the resource > > validation of ranges also not a problem? > > > > Your benchmarks are specific to this operation. Have you tried this > > patch your example 10K task / 20 node cluster to identify if the validation > > is still a bottleneck? > > > > If we run all the benchmarks in the test suite, are all of them > > positively impacted? Or are some negatively? > > > > If this is purely a win, then we can consider this approach as a > > temporary fix assuming the following: > > - Simple patch (yes! :-D ) > > - A JIRA to follow up with changing the stored data-type so we don't > > treat this as a permanent solution
Hi, Joris, thank you so much for your comment! Understand your consideration and totally agree that refactoring the resource definition should be the permanent solution. I have made a test by running 10K tasks in 20 nodes and didn't found performance bottleneck caused by resource validation. Mesos allocator now works very efficiently in that scale. And I will make further tests to evalute the impact on other test cases to see whether there is new performance regression induced. BTW, is there any approach to analyze the benchmark result(time consumption) automatically or I need to compare the difference manually? - Yanyan ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/48593/#review137281 ----------------------------------------------------------- On June 13, 2016, 12:03 p.m., Yanyan Hu wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/48593/ > ----------------------------------------------------------- > > (Updated June 13, 2016, 12:03 p.m.) > > > Review request for mesos, Guangya Liu and Joris Van Remoortere. > > > Bugs: MESOS-5425 > https://issues.apache.org/jira/browse/MESOS-5425 > > > Repository: mesos > > > Description > ------- > > This patch reimplement Ranges subtraction using > IntervalSet data type: Ranges values will be > converted to IntervalSet values for subtraction > and the result will be converted back to Ranges > after subtraction is done. This change is for > fixing jira MESOS-5425. > > > Diffs > ----- > > src/common/values.cpp 587cb68551d438621e215953e89818b623b7f71b > > Diff: https://reviews.apache.org/r/48593/diff/ > > > Testing > ------- > > make check > > > Thanks, > > Yanyan Hu > >
