Perf of Missing Values in Julia 0.7

missing_values.jl

using BenchmarkTools

n = 10_000_000

x = rand(n)
y = [ifelse(iseven(i), missing, x[i]) for i in 1:n]

sum(x)
sum(y)

@benchmark sum($x)
@benchmark sum($y)

julia> @benchmark sum($x)
BenchmarkTools.Trial:
memory estimate: 0 bytes
allocs estimate: 0

minimum time: 4.633 ms (0.00% GC)
median time: 5.285 ms (0.00% GC)
mean time: 5.444 ms (0.00% GC)
maximum time: 10.290 ms (0.00% GC)

samples: 916
evals/sample: 1

julia> @benchmark sum($y)
BenchmarkTools.Trial:
memory estimate: 0 bytes
allocs estimate: 0

minimum time: 10.661 ms (0.00% GC)
median time: 11.276 ms (0.00% GC)
mean time: 11.606 ms (0.00% GC)
maximum time: 22.728 ms (0.00% GC)

samples: 430
evals/sample: 1

library("microbenchmark")

n <- 10000000

x <- runif(n)

sum(x)

microbenchmark(sum(x))

> microbenchmark(sum(x))
Unit: milliseconds
   expr      min       lq     mean   median       uq      max neval
 sum(x) 9.147777 9.415088 9.938853 9.656602 10.27017 12.09193   100

It seems like comparing with R where there are NAs present would make sense, no?

Almost the same benchmarks appear in First-Class Statistical Missing Values Support in Julia 0.7 (do a page seach for "always allow"), but an R example with NAs is included. The result is the same as here: the R example without NAs is 2x slower than the Julia example. The R example with NAs is 2x again slower.