What's not threadsafe?

Hi all,

I am looking for a list of things that are not safe to do inside a Threads.@threads loop. I am aware that this is as high-level as it gets, and I am sorry about that.
For example, I realized myself that I cannot println from a thread (differently to doing println from an addprocs-ed worker process) before I found this issue on DiffEq. I’d love to have a list of things that don’t work - maybe here, maybe in the manual, maybe as an issue? Thanks.

One high level heuristic is that anything that does IO will not work, ie any disk reads/writes, network reads/writes or console reads/writes. Also, the GC is not threadsafe yet, so you should not allocate inside a @threads loop.

Is that really true? I think allocating inside @threads is fine. Do you have an issue or an example when this fail?

At least that doesn’t match up with what was said in the recent shared memory parallelism talk, https://youtu.be/YdiZa0Y3F3c?t=3m34s.

I think I’ve read the same somewhere - but are you sure this applies to allocations? I’d have thought that this just means you can’t invoke the gc explicitely :wink:

gc is thread safe.

No allocations at all would be quite a problem for me.

No. Both allocation and collection are thread safe.

BTW Core.println is thread-safe.

oh really? maybe I was using info. but i’m pretty sure that that was what killed this particular attempt.

Base.println is not. (At least that is my understanding.)

Regexes are not thread-safe (StrRegex.jl is though! :grinning:)
BigFloats are not completely thread-safe (because of global state for the precision and rounding mode)
(printing any floating point (Float16, Float32, Float64, BigFloat, Dec32, Dec64, Dec128, etc.) or converting it to a string used to be not thread-safe, but I believe that’s fixed now (converting or printing to an IOBuffer at least)
BigInts also used to have the problems with converting to string and printing, but I also think that’s fixed as well.

Those are the ones that I’ve run into in Base.
I’m sure there are a lot more in packages, any time there are some settable values (like the default precision or rounding mode) that are stored in global variables, they will not be thread-safe.