When benchmarking some code for EconomicScenarioGenerators.jl, I encountered this issue that I've only encountered while benchmarking. When using a shared Random.MersenneTwister
, randn
will hit a complex number and error. Here's a minimal replicating example:
function g()
v = zeros(100000)
Threads.@threads for i = 1:length(v)
v[i] =randn(Random.MersenneTwister(1))
end
v
end
function g(RNG)
v = zeros(100000)
Threads.@threads for i = 1:length(v)
v[i] =randn(RNG)
end
v
end
# these do not error
@btime g()
g()
g(Random.MersenneTwister(1))
# this errors
@btime g(Random.MersenneTwister(1))
Is this indeed an issue worth filing? If so, with Julia or with BenchmarkTools?
Full stacktrace on Mac M1, Julia 1.8 RC3
julia> @btime g(Random.MersenneTwister(1))
ERROR: TaskFailedException
Stacktrace:
[1] wait
@ ./task.jl:345 [inlined]
[2] threading_run(fun::var"#128#threadsfor_fun#31"{var"#128#threadsfor_fun#30#32"{MersenneTwister, Vector{Float64}, UnitRange{Int64}}}, static::Bool)
@ Base.Threads ./threadingconstructs.jl:38
[3] macro expansion
@ ./threadingconstructs.jl:89 [inlined]
[4] g
@ ./Untitled-1:13 [inlined]
[5] var"##core#393"()
@ Main ~/.julia/packages/BenchmarkTools/7xSXH/src/execution.jl:489
[6] var"##sample#394"(::Tuple{}, __params::BenchmarkTools.Parameters)
@ Main ~/.julia/packages/BenchmarkTools/7xSXH/src/execution.jl:497
[7] _lineartrial(b::BenchmarkTools.Benchmark, p::BenchmarkTools.Parameters; maxevals::Int64, kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
@ BenchmarkTools ~/.julia/packages/BenchmarkTools/7xSXH/src/execution.jl:161
[8] _lineartrial(b::BenchmarkTools.Benchmark, p::BenchmarkTools.Parameters)
@ BenchmarkTools ~/.julia/packages/BenchmarkTools/7xSXH/src/execution.jl:152
[9] #invokelatest#2
@ ./essentials.jl:729 [inlined]
[10] invokelatest
@ ./essentials.jl:726 [inlined]
[11] #lineartrial#46
@ ~/.julia/packages/BenchmarkTools/7xSXH/src/execution.jl:35 [inlined]
[12] lineartrial
@ ~/.julia/packages/BenchmarkTools/7xSXH/src/execution.jl:35 [inlined]
[13] tune!(b::BenchmarkTools.Benchmark, p::BenchmarkTools.Parameters; progressid::Nothing, nleaves::Float64, ndone::Float64, verbose::Bool, pad::String, kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
@ BenchmarkTools ~/.julia/packages/BenchmarkTools/7xSXH/src/execution.jl:251
[14] tune! (repeats 2 times)
@ ~/.julia/packages/BenchmarkTools/7xSXH/src/execution.jl:247 [inlined]
[15] top-level scope
@ ~/.julia/packages/BenchmarkTools/7xSXH/src/execution.jl:576
nested task error: DomainError with -1.0:
log will only return a complex result if called with a complex argument. Try log(Complex(x)).
Stacktrace:
[1] throw_complex_domainerror(f::Symbol, x::Float64)
@ Base.Math ./math.jl:33
[2] _log(x::Float64, base::Val{:ℯ}, func::Symbol)
@ Base.Math ./special/log.jl:292
[3] log
@ ./special/log.jl:257 [inlined]
[4] randn_unlikely(rng::MersenneTwister, idx::Int64, rabs::Int64, x::Float64)
@ Random ~/prog/julia-dd/usr/share/julia/stdlib/v1.8/Random/src/normal.jl:73
[5] randn
@ ~/prog/julia-dd/usr/share/julia/stdlib/v1.8/Random/src/normal.jl:54 [inlined]
[6] macro expansion
@ ./Untitled-1:14 [inlined]
[7] (::var"#128#threadsfor_fun#31"{var"#128#threadsfor_fun#30#32"{MersenneTwister, Vector{Float64}, UnitRange{Int64}}})(tid::Int64; onethread::Bool)
@ Main ./threadingconstructs.jl:84
[8] #128#threadsfor_fun
@ ./threadingconstructs.jl:51 [inlined]
[9] (::Base.Threads.var"#1#2"{var"#128#threadsfor_fun#31"{var"#128#threadsfor_fun#30#32"{MersenneTwister, Vector{Float64}, UnitRange{Int64}}}, Int64})()
@ Base.Threads ./threadingconstructs.jl:30
FWIW I hit the same thing (error only while benchmarking) with ThreadsX, but the above just uses Base to narrow in on the issue.
You should probably use the setup
argument or interpolate the rng into @btime
. It is creating too many rng's I think. Not sure who to blame yet, but I would expect you to have the same issue with @belapsed
and @benchmark
unless you setup the RNG first.
Also might want to try the newer RNG algorithm. MT has a lot of bad properties in general.
I’d assume this is an M1 Mac bug
@Michael Fiano what would you recommend? I was using MT because I thought it was the new thread-safe option
MT is the old thread-unsafe RNG
The new one is thread-local Xoshiro256++
Thanks. I was using MT based on the recommendations on the Julia blog.
Random.Xoshiro
did not errorAh yes, a good bit has changed in the last three years on this front
The new one probably could have been better advertized
@Alec The blogpost you've linked only talked about rand()
, i.e. random calls that use the thread/task specific default RNG. It does not imply that sharing MersenneTwister
instances is thread safe.
Since that blogpost was published, the default RNG was changed to Xoshiro
, due to being smaller memory wise and faster. I'm not sure if sharing instances of that is thread safe.
To the original question, it is forbidden to share objects (instances of Random.MersenneTwister or Xoshiro, for example) with mutable state between threads. This will create a data-race, and will lead to undefined behavior
The current (1.7.3) docs say:
In a multi-threaded program, you should generally use different RNG objects from different threads or tasks in order to be thread-safe. However, the default RNG is thread-safe as of Julia 1.3 (using a per-thread RNG up to version 1.6, and per-task thereafter).
Would the following clarify the situation? I current read that and think that since 1.3, the default RNG is thread safe, and now that Xoshiro is the default, it's thread-safe.
In a multi-threaded program, you should generally use different RNG objects from different threads or tasks in order to be thread-safe. However, the default global RNG is thread-safe as of Julia 1.3 (using a per-thread RNG up to version 1.6, and per-task thereafter). The global RNG refers to per-thread RNG instances; If you instantiate a non-global RNG, you should create a new RNG instance for each thread to avoid possible data-races.
julia> using Random
julia> Random.default_rng()
TaskLocalRNG()
This is a special object. The core RNG isn't thread safe.
That seems worth explicitly mentioning, since "the default" seems bound to be misinterpreted as the core/"default" RNG.
However, the default global RNG (obtained via
Random.default_rng()
) is thread-safe as of Julia 1.3
Last updated: Nov 06 2024 at 04:40 UTC