Frequently asked questions
Can I change the number of execution threads without restarting julia?
The number of execution threads is specified at the time julia is started (by command line option -t/--threads or JULIA_NUM_THREADS environment variable). Thus, to benchmark a generic Julia program with a different number of threads, you would need to start a new julia process each time.
Some libraries such as Transducers.jl, FLoops.jl, ThreadsX.jl, and Parallelism.jl support basesize option (see, e.g., Transducers.foldxt). It is used for specifying the size of the chunk of input collection processed by one task. Thus, to simulate running code with N threads, you can pass
basesize = length(input_collection) ÷ Nto run a multi-threaded function (provided that N ≤
Threads.nthreads() and there is no other function spawning tasks). For example:
julia> using ThreadsX, BenchmarkTools
julia> sum_nthreads(f, xs, N) = ThreadsX.sum(f, xs; basesize = length(xs) ÷ N);
julia> @btime sum_nthreads(sin, 1:1_000_000, 1);
16.570 ms (5 allocations: 336 bytes)
julia> @btime sum_nthreads(sin, 1:1_000_000, 2);
8.318 ms (46 allocations: 2.56 KiB)
julia> @btime sum_nthreads(sin, 1:1_000_000, 4);
4.403 ms (128 allocations: 7.03 KiB)Note that this trick cannot be used for experimenting the effect of the number of threads to the load-balancing of multi-threaded code since load-balancing requires starting more than Threads.nthreads() tasks.
Should I use multi-threading? Or should I use multi-processing?
Julia supports threading-based (via Base.Threads) and process-based (via Distributed.jl) parallelism paradigms. Each paradigm has pros and cons. Choosing the best option requires understanding what your program does.
Multi-threading is better for processing complex and large objects whose serialization become bottleneck in multi-processing -based parallelism. Note that DistributedArrays.jl can be used to reduce serialization overhead in multi-processing.
If your code allocates many intermediate objects, multi-processing -based frameworks such as Distributed.jl standard library, Dagger.jl, or MPI.jl are better option. This is because julia's memory management system (garbage collection; GC) can be a bottleneck for scaling such type of code to many execution threads.
To make it easy to balance with these trade-offs, it is recommended to use a high-level of abstraction such as data parallelism that helps you switch underlying execution mechanisms. For example JuliaFolds packages such as Folds.jl, FLoops.jl, and Transducers.jl have executor argument to easily switch thread-based and process-based execution mechanisms.
Why is the approach using state[threadid()] not mentioned?
See: What is the difference of @reduce and @init to the approach using state[threadid()]? · FAQ · FLoops
