ThreadTools
This package implements some utilities for using threads in Julia v1.3+
The utilities provided are:
@spawnatmost n for i in ...Spawn at most n threads to carry out for-loop
tmap(f, args...)
tmap(f, nthreads::Int, args...)
tmap1(f, args...)
tmap1(f, nthreads::Int,args...)Threaded map. The optional argument nthreads limits the number of threads used in parallel.
tmap1 is the same as tmap, but falls back to a regular map if julia only has access to one thread.
The tools to limit the number of threads used are useful when doing, e.g., threaded IO, where the disk might get overloaded if you try to access it from too many threads at the same time.
Examples
using ThreadTools
julia> l = SpinLock();
julia> times = [];
julia> @spawnatmost 3 for i = 1:10 # This will use only three parallel threads, even if more are avilable
lock(l) do # We protect the access to the array using a lock
push!(times, time())
end
println(i)
sleep(1)
end
julia> round.(diff(times), digits=2)
9-element Array{Float64,1}:
0.0
0.0
1.0
0.0
0.0
1.0
0.0
0.0
1.0
julia> tmap(_->threadid(), 1:5) # A threaded version of map
5-element Array{Int64,1}:
2
6
3
4
5
julia> times = tmap(_->(t=time();sleep(0.3);t), 3, 1:10); # The second argument limits the number of threads used
julia> round.(diff(times), digits=2)
9-element Array{Float64,1}:
0.0
0.0
0.3
0.0
0.0
0.3
0.0
0.0
0.3Benchmark
All benchmarks are done on a machine with 6 physical cores. The following function will be used for benchmarking
function fib(n)
if n <= 1 return 1 end
return fib(n - 1) + fib(n - 2)
endLight workload
fib(20) 720 times
workload = fill(20, factorial(6))
times = map(1:6) do nt
nt == 1 && return @belapsed map(fib, workload)
@belapsed tmap(fib, $nt, workload)
end
t720 = @belapsed tmap(fib, workload)
plot(1:6, times, xlabel="Number of threads", ylabel="Time [s]")
hline!([t720], label="720 tasks", ylims=(0,Inf))
Conclusion: for light loads, the overhead is large. You may want to consider KissThreading.jl for a more advanced approach.
Heavy workload
fib(30) 720 times
workload = fill(30, factorial(6))
times = map(1:6) do nt
nt == 1 && return @belapsed map(fib, workload)
@belapsed tmap(fib, $nt, workload)
end
t720 = @belapsed tmap(fib, workload)
plot(1:6, times, xlabel="Number of threads", ylabel="Time [s]")
hline!([t720], label="720 tasks", ylims=(0,Inf))
Conclusion: tmap is effective when the computational load is heavy.