Switch to Gadfly, set colors based on library

[chriselrod]

Sample plot:

Resolves #28.

[DilumAluthge]

Ooof, what happened to Gaius here?

[chriselrod]

Yeah, this definitely looks like a regression. The actual numbers

julia>  rb = runbench(sizes=logspace(20,300,21)); plot(rb)

julia> rb
Bennchmark Result of Matrix{Float64}, threaded = true
21×7 DataFrame
 Row │ Sizes  BLIS        Gaius      MKL        Octavian   OpenBLAS  Tullio
     │ Int64  Float64     Float64    Float64    Float64    Float64   Float64
─────┼─────────────────────────────────────────────────────────────────────────
   1 │    20    0.377902  60.2249      35.5883    65.3644   22.6062   59.9568
   2 │    23    0.571999  70.8202      46.9537    75.5972   20.4935   71.7154
   3 │    26    0.827262  61.0035      42.6486    63.0687   27.9405   62.4411
   4 │    30    1.26692   71.122       52.0432    72.3764   30.6644   73.3924
   5 │    34    1.77145   71.2416      57.8766    72.8999   35.6427   72.2102
   6 │    39    2.65987   75.8313      66.0641    78.7612   30.137    78.0256
   7 │    45    4.05288    0.680721    75.048     87.7937   36.4148   85.4787
   8 │    52    4.69162    1.02951     68.5057   145.904    45.9023   86.3586
   9 │    59    6.15838    1.49911     97.1748   173.307    39.6217   70.5325
  10 │    68    8.09943    2.38282    171.108    249.538    60.1841    6.91789
  11 │    77   11.4472    10.0319     210.696    285.523    71.4729   10.6282
  12 │    89   14.4787    15.7012     263.015    313.818   100.266    10.3768
  13 │   102   21.6372    17.4339     358.347    412.12    138.965    17.2046
  14 │   116   29.3559    20.6917     523.018    493.268   169.663    25.3127
  15 │   133   39.4503    28.8287     628.207    603.743   221.571    24.6444
  16 │   152   62.5768    25.7061     860.808   1087.99    215.554    43.8051
  17 │   175   77.6305    30.6713     845.727    977.631   261.051    48.4
  18 │   200  103.411     29.363     1319.37    1430.1     317.058    68.9239
  19 │   229  156.828     36.9907     842.707   1062.98    452.563    97.9942
  20 │   262  154.415     48.7173    1067.03    1288.35    520.474    97.1263
  21 │   300  217.563     44.7269    1210.57    1144.31    543.396   139.214

EDIT: Oh, this might be breaking, because it changed what the BenchmarkResult struct looks like.

[stillyslalom]

Lighter colors (particularly Tullio's) are hard to make out against the white background

[chriselrod]

On the computer that ran the benchmarks from the top of this thread:

julia> versioninfo()
Julia Version 1.6.0-beta1.1
Commit 8240225853* (2021-01-11 19:36 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i9-7980XE CPU @ 2.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-11.0.0 (ORCJIT, skylake-avx512)
Environment:
  JULIA_NUM_THREADS = 36
  JULIA_PATH = /home/chriselrod/Documents/languages/julia/usr

julia> using BenchmarkTools

julia> @benchmark wait(Threads.@spawn nothing)
BenchmarkTools.Trial:
  memory estimate:  448 bytes
  allocs estimate:  5
  --------------
  minimum time:     65.176 μs (0.00% GC)
  median time:      73.731 μs (0.00% GC)
  mean time:        73.969 μs (0.00% GC)
  maximum time:     111.876 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     1

This was in a fresh Julia session, started with --startup=no.
This is a 7980XE, and got roughly the same results on 1.6-beta and on master.

My 10980XE -- basically the same CPU -- got this a few days ago (copy/pasted from JuliaLinearAlgebra/Octavian.jl#24 (comment)):

julia> @benchmark wait(Threads.@spawn nothing) # just spawning a thread and waiting on it
BenchmarkTools.Trial:
  memory estimate:  443 bytes
  allocs estimate:  4
  --------------
  minimum time:     1.811 μs (0.00% GC)
  median time:      7.741 μs (0.00% GC)
  mean time:        9.088 μs (0.00% GC)
  maximum time:     57.478 μs (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     7

Threads.@spawn explains the bad performance of both Tullio and Gaius. The 10980XE is unavailable for the next 36 hours or so, but I can look more closely later. I tried on a a 7900X and got 2.4/40.3/37.4 min/median/mean, which is sort of between.
The 7900X also has 4 allocs, 443 bytes.

Maybe it's related to OS/kernel parameters. But I have a hard time believing I wouldn't have noticed this earlier, unless it's a recent regression somewhere.