Yup, got it. bisecting now.
On Tuesday, June 30, 2015 at 6:02:06 PM UTC-5, andrew cooke wrote: > > bisect is log(n)... > > On Tuesday, 30 June 2015 19:36:03 UTC-3, Seth wrote: >> >> I can't tell what changed to cause this. The versions are 35 days apart. >> >> On Tuesday, June 30, 2015 at 5:10:52 PM UTC-5, Kristoffer Carlsson wrote: >>> >>> If it is a specific commit that introduced the regression it should go >>> fairly quick to git bisect it down. Maybe you can try that? >>> >>> On Tuesday, June 30, 2015 at 11:07:35 PM UTC+2, Seth wrote: >>>> >>>> Please replace "vertex" with "edge" below. Sorry about that. >>>> >>>> On Tuesday, June 30, 2015 at 4:03:15 PM UTC-5, Seth wrote: >>>>> >>>>> I'm seeing a fairly significant performance regression between Julia >>>>> Version 0.4.0-dev+5008 and Julia Version 0.4.0-dev+5712 ,and I don't >>>>> know how to go about identifying what's wrong. >>>>> >>>>> I'm timing betweenness_centrality in LightGraphs.jl. It used to be >>>>> among the fastest implementations out there, but it's slipping. >>>>> >>>>> Here are some samples: >>>>> >>>>> for a 5k-node, 50k-vertex graph: >>>>> >>>>> 5008: 12.581 seconds (75506 k allocations: 5047 MB, 5.26% gc time) >>>>> 5712: 19.291 seconds (75507 k allocations: 5047 MB) >>>>> >>>>> for a 10k-node, 100k-vertex graph: >>>>> >>>>> 5008: 91.238 seconds (380 M allocations: 27234 MB, 6.21% gc time) >>>>> 5712: 100.042 seconds (380 M allocations: 27235 MB, 5.97% gc time) >>>>> >>>>> for the benchmark (a 39796-node, 301498-edge graph: >>>>> >>>>> 5008: 1238 seconds (3769 M allocations: 272 GB, 18.38% gc time) >>>>> 5712: 1515 seconds (3769 M allocations: 272 GB, 16.63% gc time) >>>>> >>>>> Any ideas on how to figure out what's going on here? Profile shows the >>>>> same lines in the same order (though for some reason 5712 is showing a >>>>> lot >>>>> more internal profiling). >>>>> >>>>
