Area LibGEOS Polygon problems

[skygering]

Area LibGEOS Polygon problems

[rafaqz]

It could be type instability too

Or an inneficient GeoInterface path

(nope its just LibGEOS)

[evetion]

This was runtime dispatch on GeoInterface.x and y. Solved by JuliaGeo/LibGEOS.jl#191, which brings the benchmark to 4μs for me (still slow, but it's copying points from C, so expected).

[rafaqz]

That's not time for copying points from C - it's mostly for allocating arrays in Julia, the yellow Array

[evetion]

That's not time for copying points from C - it's mostly for allocating arrays in Julia, the yellow Array

That picture was the before state. The afterward profile is in the other issue in LibGEOS.

[rafaqz]

Ok right, I think there are still a bunch of allocations happening though when each point is loaded.

I can get it 7x faster changing area in the benchmark to use getpoint with getgeom(linestring) implemented in JuliaGeo/LibGEOS.jl#193.

See #36 for faster area and benchmarks

[rafaqz]

Also see:
JuliaGeo/LibGEOS.jl#196

Turns out most of the time is making clones. Without them we are down to:

julia> @benchmark GO.area($pLG)
BenchmarkTools.Trial: 10000 samples with 337 evaluations.
 Range (min … max):  256.193 ns …  10.465 μs  ┊ GC (min … max): 0.00% … 96.32%
 Time  (median):     260.564 ns               ┊ GC (median):    0.00%
 Time  (mean ± σ):   277.483 ns ± 167.507 ns  ┊ GC (mean ± σ):  1.02% ±  1.67%

  ▅█▄▅▄▄▃▂▁  ▁        ▂▁ ▂  ▁▃  ▂▁                              ▂
  ████████████▇▇██▇█▇█████▇▆██████▇▇█▇▅▅▇▇▃▅▄▇▃▃▄▃▆▃▁▁▃▇▃▁▁▄▄▅▅ █
  256 ns        Histogram: log(frequency) by time        418 ns <

 Memory estimate: 48 bytes, allocs estimate: 2.

[rafaqz]

With some fixes to context:

julia> @benchmark GO.area($pLG)
BenchmarkTools.Trial: 10000 samples with 535 evaluations.
 Range (min … max):  210.733 ns …   6.821 μs  ┊ GC (min … max): 0.00% … 96.09%
 Time  (median):     213.314 ns               ┊ GC (median):    0.00%
 Time  (mean ± σ):   220.193 ns ± 135.810 ns  ┊ GC (mean ± σ):  1.36% ±  2.14%

  ▄██▄▂▃▄▁  ▁▁                                     ▂      ▁     ▂
  ████████▇████▇▅▆▅▆▆▅▆▅▅▄▅▆▅▅▃▃▅▅▁▁▃▅█▇▄▁▃▅█▇▄▁▅▃██▇▁▄▅▄██▅▁▁▄ █
  211 ns        Histogram: log(frequency) by time        284 ns <

 Memory estimate: 48 bytes, allocs estimate: 2.

[rafaqz]

And fixing how we get points a bit more:

julia> @benchmark GO.area($pLG)
BenchmarkTools.Trial: 10000 samples with 870 evaluations.
 Range (min … max):  128.115 ns …   4.316 μs  ┊ GC (min … max): 0.00% … 95.92%
 Time  (median):     130.037 ns               ┊ GC (median):    0.00%
 Time  (mean ± σ):   138.450 ns ± 117.333 ns  ┊ GC (mean ± σ):  2.64% ±  3.03%

  ▅█▆▃▅▄ ▃▃▁▁▃▃                                  ▁              ▁
  █████████████▇▆▆▆▆▆▇▇▇█▇▇▆▇▇▇▇▇▇▅▅▅▄▅▇▆▅▅▅██▅▅▆███▅▆▆▇▇▆▃▃▂▃▆ █
  128 ns        Histogram: log(frequency) by time        185 ns <

 Memory estimate: 64 bytes, allocs estimate: 3.

So under 4x using LibGEOS directly is possible.