Lookup in Dict{Int,Float64} allocates

General Usage Performance
,
1

I was surprised that the following code reports an allocation when looking up a Float64 value from a Dict{Int64,Float64}

using BenchmarkTools

n = 1_000_000
d = Dict(i=>rand() for i in 1:n);
@btime d[42]
# 19.346 ns (1 allocation: 16 bytes)

even more strangely, if I make a Dict{Int64,String} instead (like was done here), I get 0 allocations

using BenchmarkTools
using Random

n = 1_000_000
d = Dict(i=>randstring() for i in 1:n);
@btime d[42]
# 19.346 ns (1 allocation: 16 bytes)

It seems bizzare that the allocations would depend on the value type of the Dict!

This is not actually a performance issue in my code, because I only have a small number of Dict lookups. I just stumbled across these allocations while profiling and was curious why they happen at all. Does anyone know why?

2

Don’t benchmark in global scope like that. See some of the top tips in Performance Tips · The Julia Language.

3

To elaborate on what Kristoffer said, you can either interpolate d from global scope:

julia> @btime $(d)[42]
  4.651 ns (0 allocations: 0 bytes)
0.852560373628357

or create a new local d using BenchmarkTools tools:

julia> @btime d[42] setup=(d = Dict(i=>rand() for i in 1:1_000_000))
  4.720 ns (0 allocations: 0 bytes)
0.20694337049842937
4 Likes