Multi-Threading

Visit this blog post for a presentation of Julia multi-threading features.

Starting Julia with multiple threads

By default, Julia starts up with a single thread of execution. This can be verified by using the command Threads.nthreads():

julia> Threads.nthreads()
1

The number of execution threads is controlled either by using the -t/--threads command line argument or by using the JULIA_NUM_THREADS environment variable. When both are specified, then -t/--threads takes precedence.

The number of threads can either be specified as an integer (--threads=4) or as auto (--threads=auto), where auto tries to infer a useful default number of threads to use (see Command-line Options for more details).

Lets start Julia with 4 threads:

$ julia --threads 4

Let's verify there are 4 threads at our disposal.

julia> Threads.nthreads()
4

But we are currently on the master thread. To check, we use the function Threads.threadid

julia> Threads.threadid()
1

export JULIA_NUM_THREADS=4

set JULIA_NUM_THREADS=4

$env:JULIA_NUM_THREADS=4

Multiple GC Threads

The Garbage Collector (GC) can use multiple threads. The amount used is either half the number of compute worker threads or configured by either the --gcthreads command line argument or by using the JULIA_NUM_GC_THREADS environment variable.

Threadpools

When a program's threads are busy with many tasks to run, tasks may experience delays which may negatively affect the responsiveness and interactivity of the program. To address this, you can specify that a task is interactive when you Threads.@spawn it:

using Base.Threads
@spawn :interactive f()

Interactive tasks should avoid performing high latency operations, and if they are long duration tasks, should yield frequently.

Julia may be started with one or more threads reserved to run interactive tasks:

$ julia --threads 3,1

The environment variable JULIA_NUM_THREADS can also be used similarly:

export JULIA_NUM_THREADS=3,1

This starts Julia with 3 threads in the :default threadpool and 1 thread in the :interactive threadpool:

julia> using Base.Threads

julia> nthreadpools()
2

julia> threadpool() # the main thread is in the interactive thread pool
:interactive

julia> nthreads(:default)
3

julia> nthreads(:interactive)
1

julia> nthreads()
3

Either or both numbers can be replaced with the word auto, which causes Julia to choose a reasonable default.

The @threads Macro

Let's work a simple example using our native threads. Let us create an array of zeros:

julia> a = zeros(10)
10-element Vector{Float64}:
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0
 0.0

Let us operate on this array simultaneously using 4 threads. We'll have each thread write its thread ID into each location.

Julia supports parallel loops using the Threads.@threads macro. This macro is affixed in front of a for loop to indicate to Julia that the loop is a multi-threaded region:

julia> Threads.@threads for i = 1:10
           a[i] = Threads.threadid()
       end

The iteration space is split among the threads, after which each thread writes its thread ID to its assigned locations:

julia> a
10-element Vector{Float64}:
 1.0
 1.0
 1.0
 2.0
 2.0
 2.0
 3.0
 3.0
 4.0
 4.0

Note that Threads.@threads does not have an optional reduction parameter like @distributed.

Using @threads without data-races

The concept of a data-race is elaborated on in "Communication and data races between threads". For now, just known that a data race can result in incorrect results and dangerous errors.

Lets say we want to make the function sum_single below multithreaded.

julia> function sum_single(a)
           s = 0
           for i in a
               s += i
           end
           s
       end
sum_single (generic function with 1 method)

julia> sum_single(1:1_000_000)
500000500000

Simply adding @threads exposes a data race with multiple threads reading and writing s at the same time.

julia> function sum_multi_bad(a)
           s = 0
           Threads.@threads for i in a
               s += i
           end
           s
       end
sum_multi_bad (generic function with 1 method)

julia> sum_multi_bad(1:1_000_000)
70140554652

Note that the result is not 500000500000 as it should be, and will most likely change each evaluation.

To fix this, buffers that are specific to the task may be used to segment the sum into chunks that are race-free. Here sum_single is reused, with its own internal buffer s. The input vector a is split into at most nthreads() chunks for parallel work. We then use Threads.@spawn to create tasks that individually sum each chunk. Finally, we sum the results from each task using sum_single again:

julia> function sum_multi_good(a)
           chunks = Iterators.partition(a, cld(length(a), Threads.nthreads()))
           tasks = map(chunks) do chunk
               Threads.@spawn sum_single(chunk)
           end
           chunk_sums = fetch.(tasks)
           return sum_single(chunk_sums)
       end
sum_multi_good (generic function with 1 method)

julia> sum_multi_good(1:1_000_000)
500000500000

Another option is the use of atomic operations on variables shared across tasks/threads, which may be more performant depending on the characteristics of the operations.

Communication and data-races between threads

Although Julia's threads can communicate through shared memory, it is notoriously difficult to write correct and data-race free multi-threaded code. Julia's Channels are thread-safe and may be used to communicate safely. There are also sections below that explain how to use locks and atomics to avoid data-races.

In certain cases, Julia is able to detect a detect safety violations, in particular in regards to deadlocks or other known-unsafe operations such as yielding to the currently running task. In these cases, a ConcurrencyViolationError is thrown.

Data-race freedom

You are entirely responsible for ensuring that your program is data-race free, and nothing promised here can be assumed if you do not observe that requirement. The observed results may be highly unintuitive.

If data-races are introduced, Julia is not memory safe. Be very careful about reading any data if another thread might write to it, as it could result in segmentation faults or worse. Below are a couple of unsafe ways to access global variables from different threads:

Thread 1:
global b = false
global a = rand()
global b = true

Thread 2:
while !b; end
bad_read1(a) # it is NOT safe to access `a` here!

Thread 3:
while !@isdefined(a); end
bad_read2(a) # it is NOT safe to access `a` here

Using locks to avoid data-races

An important tool to avoid data-races, and thereby write thread-safe code, is the concept of a "lock". A lock can be locked and unlocked. If a thread has locked a lock, and not unlocked it, it is said to "hold" the lock. If there is only one lock, and we write code the requires holding the lock to access some data, we can ensure that multiple threads will never access the same data simultaneously. Note that the link between a lock and a variable is made by the programmer, and not the program.

For example, we can create a lock my_lock, and lock it while we mutate a variable my_variable. This is done most simply with the @lock macro:

julia> my_lock = ReentrantLock();

julia> my_variable = [1, 2, 3];

julia> @lock my_lock my_variable[1] = 100
100

By using a similar pattern with the same lock and variable, but on another thread, the operations are free from data-races.

We could have performed the operation above with the functional version of lock, in the following two ways:

julia> lock(my_lock) do
           my_variable[1] = 100
       end
100

julia> begin
           lock(my_lock)
           try
               my_variable[1] = 100
           finally
               unlock(my_lock)
           end
       end
100

All three options are equivalent. Note how the final version requires an explicit try-block to ensure that the lock is always unlocked, whereas the first two version do this internally. One should always use the lock pattern above when changing data (such as assigning to a global or closure variable) accessed by other threads. Failing to do this could have unforeseen and serious consequences.

Atomic Operations

Julia supports accessing and modifying values atomically, that is, in a thread-safe way to avoid race conditions. A value (which must be of a primitive type) can be wrapped as Threads.Atomic to indicate it must be accessed in this way. Here we can see an example:

julia> i = Threads.Atomic{Int}(0);

julia> ids = zeros(4);

julia> old_is = zeros(4);

julia> Threads.@threads for id in 1:4
           old_is[id] = Threads.atomic_add!(i, id)
           ids[id] = id
       end

julia> old_is
4-element Vector{Float64}:
 0.0
 1.0
 7.0
 3.0

julia> i[]
 10

julia> ids
4-element Vector{Float64}:
 1.0
 2.0
 3.0
 4.0

Had we tried to do the addition without the atomic tag, we might have gotten the wrong answer due to a race condition. An example of what would happen if we didn't avoid the race:

julia> using Base.Threads

julia> Threads.nthreads()
4

julia> acc = Ref(0)
Base.RefValue{Int64}(0)

julia> @threads for i in 1:1000
          acc[] += 1
       end

julia> acc[]
926

julia> acc = Atomic{Int64}(0)
Atomic{Int64}(0)

julia> @threads for i in 1:1000
          atomic_add!(acc, 1)
       end

julia> acc[]
1000

Per-field atomics

We can also use atomics on a more granular level using the @atomic, @atomicswap, @atomicreplace macros, and @atomiconce macros.

Specific details of the memory model and other details of the design are written in the Julia Atomics Manifesto, which will later be published formally.

Any field in a struct declaration can be decorated with @atomic, and then any write must be marked with @atomic also, and must use one of the defined atomic orderings (:monotonic, :acquire, :release, :acquire_release, or :sequentially_consistent). Any read of an atomic field can also be annotated with an atomic ordering constraint, or will be done with monotonic (relaxed) ordering if unspecified.

Side effects and mutable function arguments

When using multi-threading we have to be careful when using functions that are not pure as we might get a wrong answer. For instance functions that have a name ending with ! by convention modify their arguments and thus are not pure.

@threadcall

External libraries, such as those called via ccall, pose a problem for Julia's task-based I/O mechanism. If a C library performs a blocking operation, that prevents the Julia scheduler from executing any other tasks until the call returns. (Exceptions are calls into custom C code that call back into Julia, which may then yield, or C code that calls jl_yield(), the C equivalent of yield.)

The @threadcall macro provides a way to avoid stalling execution in such a scenario. It schedules a C function for execution in a separate thread. A threadpool with a default size of 4 is used for this. The size of the threadpool is controlled via environment variable UV_THREADPOOL_SIZE. While waiting for a free thread, and during function execution once a thread is available, the requesting task (on the main Julia event loop) yields to other tasks. Note that @threadcall does not return until the execution is complete. From a user point of view, it is therefore a blocking call like other Julia APIs.

It is very important that the called function does not call back into Julia, as it will segfault.

@threadcall may be removed/changed in future versions of Julia.

Caveats

At this time, most operations in the Julia runtime and standard libraries can be used in a thread-safe manner, if the user code is data-race free. However, in some areas work on stabilizing thread support is ongoing. Multi-threaded programming has many inherent difficulties, and if a program using threads exhibits unusual or undesirable behavior (e.g. crashes or mysterious results), thread interactions should typically be suspected first.

There are a few specific limitations and warnings to be aware of when using threads in Julia:

• Base collection types require manual locking if used simultaneously by multiple threads where at least one thread modifies the collection (common examples include push! on arrays, or inserting items into a Dict).
• The schedule used by @spawn is nondeterministic and should not be relied on.
• Compute-bound, non-memory-allocating tasks can prevent garbage collection from running in other threads that are allocating memory. In these cases it may be necessary to insert a manual call to GC.safepoint() to allow GC to run. This limitation will be removed in the future.
• Avoid running top-level operations, e.g. include, or eval of type, method, and module definitions in parallel.
• Be aware that finalizers registered by a library may break if threads are enabled. This may require some transitional work across the ecosystem before threading can be widely adopted with confidence. See the section on the safe use of finalizers for further details.

Task Migration

After a task starts running on a certain thread it may move to a different thread if the task yields.

Such tasks may have been started with @spawn or @threads, although the :static schedule option for @threads does freeze the threadid.

This means that in most cases threadid() should not be treated as constant within a task, and therefore should not be used to index into a vector of buffers or stateful objects.

Safe use of Finalizers

Because finalizers can interrupt any code, they must be very careful in how they interact with any global state. Unfortunately, the main reason that finalizers are used is to update global state (a pure function is generally rather pointless as a finalizer). This leads us to a bit of a conundrum. There are a few approaches to dealing with this problem:

1. When single-threaded, code could call the internal jl_gc_enable_finalizers C function to prevent finalizers from being scheduled inside a critical region. Internally, this is used inside some functions (such as our C locks) to prevent recursion when doing certain operations (incremental package loading, codegen, etc.). The combination of a lock and this flag can be used to make finalizers safe.

2. A second strategy, employed by Base in a couple places, is to explicitly delay a finalizer until it may be able to acquire its lock non-recursively. The following example demonstrates how this strategy could be applied to Distributed.finalize_ref:

   function finalize_ref(r::AbstractRemoteRef)
       if r.where > 0 # Check if the finalizer is already run
           if islocked(client_refs) || !trylock(client_refs)
               # delay finalizer for later if we aren't free to acquire the lock
               finalizer(finalize_ref, r)
               return nothing
           end
           try # `lock` should always be followed by `try`
               if r.where > 0 # Must check again here
                   # Do actual cleanup here
                   r.where = 0
               end
           finally
               unlock(client_refs)
           end
       end
       nothing
   end

3. A related third strategy is to use a yield-free queue. We don't currently have a lock-free queue implemented in Base, but Base.IntrusiveLinkedListSynchronized{T} is suitable. This can frequently be a good strategy to use for code with event loops. For example, this strategy is employed by Gtk.jl to manage lifetime ref-counting. In this approach, we don't do any explicit work inside the finalizer, and instead add it to a queue to run at a safer time. In fact, Julia's task scheduler already uses this, so defining the finalizer as x -> @spawn do_cleanup(x) is one example of this approach. Note however that this doesn't control which thread do_cleanup runs on, so do_cleanup would still need to acquire a lock. That doesn't need to be true if you implement your own queue, as you can explicitly only drain that queue from your thread.