Add a :dynamic scheduling option for Threads.@threads (#43919)

Co-authored-by: Julian Samaroo <[email protected]> Co-authored-by: Takafumi Arakaki <[email protected]> Co-authored-by: Valentin Churavy <[email protected]>
JuliaLang · Apr 14, 2022 · 02b3f81 · 02b3f81
1 parent b174ed8
commit 02b3f81
Show file tree

Hide file tree

Showing 3 changed files with 132 additions and 17 deletions.
diff --git a/NEWS.md b/NEWS.md
@@ -73,6 +73,9 @@ Command-line option changes
 Multi-threading changes
 -----------------------
 
+* A new `:dynamic` schedule option for `Threads.@threads` which is similar to the default behavior except iterations
+  will be scheduled dynamically to available worker threads rather than pinned to each thread. This option is more
+  composable with (possibly nested) `@spawn` and `@threads` loops ([#43919])
 
 Build system changes
 --------------------

diff --git a/base/threadingconstructs.jl b/base/threadingconstructs.jl
@@ -22,14 +22,14 @@ See also: `BLAS.get_num_threads` and `BLAS.set_num_threads` in the
 """
 nthreads() = Int(unsafe_load(cglobal(:jl_n_threads, Cint)))
 
-function threading_run(func)
+function threading_run(fun, static)
     ccall(:jl_enter_threaded_region, Cvoid, ())
     n = nthreads()
     tasks = Vector{Task}(undef, n)
     for i = 1:n
-        t = Task(func)
-        t.sticky = true
-        ccall(:jl_set_task_tid, Cint, (Any, Cint), t, i-1)
+        t = Task(() -> fun(i)) # pass in tid
+        t.sticky = static
+        static && ccall(:jl_set_task_tid, Cint, (Any, Cint), t, i-1)
         tasks[i] = t
         schedule(t)
     end
@@ -48,15 +48,14 @@ function _threadsfor(iter, lbody, schedule)
     quote
         local threadsfor_fun
         let range = $(esc(range))
-        function threadsfor_fun(onethread=false)
+        function threadsfor_fun(tid=1; onethread=false)
             r = range # Load into local variable
             lenr = length(r)
             # divide loop iterations among threads
             if onethread
                 tid = 1
                 len, rem = lenr, 0
             else
-                tid = threadid()
                 len, rem = divrem(lenr, nthreads())
             end
             # not enough iterations for all the threads?
@@ -86,15 +85,17 @@ function _threadsfor(iter, lbody, schedule)
             end
         end
         end
-        if ccall(:jl_in_threaded_region, Cint, ()) != 0
+        if $(schedule === :dynamic)
+            threading_run(threadsfor_fun, false)
+        elseif ccall(:jl_in_threaded_region, Cint, ()) != 0
             $(if schedule === :static
               :(error("`@threads :static` cannot be used concurrently or nested"))
               else
               # only use threads when called from outside @threads
-              :(threadsfor_fun(true))
+              :(threadsfor_fun(onethread = true))
               end)
         else
-            threading_run(threadsfor_fun)
+            threading_run(threadsfor_fun, true)
         end
         nothing
     end
@@ -110,15 +111,73 @@ A barrier is placed at the end of the loop which waits for all tasks to finish
 execution.
 
 The `schedule` argument can be used to request a particular scheduling policy.
-The only currently supported value is `:static`, which creates one task per thread
-and divides the iterations equally among them. Specifying `:static` is an error
-if used from inside another `@threads` loop or from a thread other than 1.
+
+Except for `:static` scheduling, how the iterations are assigned to tasks, and how the tasks
+are assigned to the worker threads is undefined. The exact assignments can be different
+for each execution. The scheduling option is a hint. The loop body code (including any code
+transitively called from it) must not make assumptions about the distribution of iterations
+to tasks or the worker thread in which they are executed. The loop body for each iteration
+must be able to make forward progress independent of other iterations and be free from data
+races. As such, synchronizations across iterations may deadlock.
+
+For example, the above conditions imply that:
+
+- The lock taken in an iteration *must* be released within the same iteration.
+- Communicating between iterations using blocking primitives like `Channel`s is incorrect.
+- Write only to locations not shared across iterations (unless a lock or atomic operation is used).
+
+
+Schedule options are:
+- `:static` creates one task per thread and divides the iterations equally among
+            them, assigning each task specifically to each thread.
+            Specifying `:static` is an error if used from inside another `@threads` loop
+            or from a thread other than 1.
+- `:dynamic` will schedule iterations dynamically to available worker threads,
+            assuming that the workload for each iteration is uniform.
+
+Without the scheduler argument, the exact scheduling is unspecified; i.e. it may be
+different across Julia releases. Currently, the behavior is dependent on the calling thread.
+The default is `:static` when called from thread 1. The loop will be executed without threading
+when called from other threads.
 
 The default schedule (used when no `schedule` argument is present) is subject to change.
 
+For example, an illustration of the different scheduling strategies where `busywait`
+is a non-yielding timed loop that runs for a number of seconds.
+
+```julia-repl
+julia> function busywait(seconds)
+            tstart = time_ns()
+            while (time_ns() - tstart) / 1e9 < seconds
+            end
+        end
+
+julia> @time begin
+            Threads.@spawn busywait(5)
+            Threads.@threads :static for i in 1:Threads.nthreads()
+                busywait(1)
+            end
+        end
+6.003001 seconds (16.33 k allocations: 899.255 KiB, 0.25% compilation time)
+
+julia> @time begin
+            Threads.@spawn busywait(5)
+            Threads.@threads :dynamic for i in 1:Threads.nthreads()
+                busywait(1)
+            end
+        end
+2.012056 seconds (16.05 k allocations: 883.919 KiB, 0.66% compilation time)
+```
+
+The `:dynamic` example takes 2 seconds since one of the non-occupied threads is able
+to run two of the 1-second iterations to complete the for loop.
+
 !!! compat "Julia 1.5"
     The `schedule` argument is available as of Julia 1.5.
 
+!!! compat "Julia 1.8"
+    The `:dynamic` option for the `schedule` argument is available as of Julia 1.8.
+
 See also: [`@spawn`](@ref Threads.@spawn), [`nthreads()`](@ref Threads.nthreads),
 [`threadid()`](@ref Threads.threadid), `pmap` in [`Distributed`](@ref man-distributed),
 `BLAS.set_num_threads` in [`LinearAlgebra`](@ref man-linalg).
@@ -133,7 +192,7 @@ macro threads(args...)
             # for now only allow quoted symbols
             sched = nothing
         end
-        if sched !== :static
+        if sched !== :static && sched !== :dynamic
             throw(ArgumentError("unsupported schedule argument in @threads"))
         end
     elseif na == 1

diff --git a/test/threads_exec.jl b/test/threads_exec.jl
@@ -722,15 +722,68 @@ let a = zeros(nthreads())
 end
 
 # static schedule
-function _atthreads_static_schedule()
+function _atthreads_static_schedule(n)
+    ids = zeros(Int, n)
+    Threads.@threads :static for i = 1:n
+        ids[i] = Threads.threadid()
+    end
+    return ids
+end
+@test _atthreads_static_schedule(nthreads()) == 1:nthreads()
+@test _atthreads_static_schedule(1) == [1;]
+@test_throws(
+    "`@threads :static` cannot be used concurrently or nested",
+    @threads(for i = 1:1; _atthreads_static_schedule(nthreads()); end),
+)
+
+# dynamic schedule
+function _atthreads_dynamic_schedule(n)
+    inc = Threads.Atomic{Int}(0)
+    flags = zeros(Int, n)
+    Threads.@threads :dynamic for i = 1:n
+        Threads.atomic_add!(inc, 1)
+        flags[i] = 1
+    end
+    return inc[], flags
+end
+@test _atthreads_dynamic_schedule(nthreads()) == (nthreads(), ones(nthreads()))
+@test _atthreads_dynamic_schedule(1) == (1, ones(1))
+@test _atthreads_dynamic_schedule(10) == (10, ones(10))
+@test _atthreads_dynamic_schedule(nthreads() * 2) == (nthreads() * 2, ones(nthreads() * 2))
+
+# nested dynamic schedule
+function _atthreads_dynamic_dynamic_schedule()
+    inc = Threads.Atomic{Int}(0)
+    Threads.@threads :dynamic for _ = 1:nthreads()
+        Threads.@threads :dynamic for _ = 1:nthreads()
+            Threads.atomic_add!(inc, 1)
+        end
+    end
+    return inc[]
+end
+@test _atthreads_dynamic_dynamic_schedule() == nthreads() * nthreads()
+
+function _atthreads_static_dynamic_schedule()
     ids = zeros(Int, nthreads())
+    inc = Threads.Atomic{Int}(0)
     Threads.@threads :static for i = 1:nthreads()
         ids[i] = Threads.threadid()
+        Threads.@threads :dynamic for _ = 1:nthreads()
+            Threads.atomic_add!(inc, 1)
+        end
     end
-    return ids
+    return ids, inc[]
+end
+@test _atthreads_static_dynamic_schedule() == (1:nthreads(), nthreads() * nthreads())
+
+# errors inside @threads :dynamic
+function _atthreads_dynamic_with_error(a)
+    Threads.@threads :dynamic for i in eachindex(a)
+        error("user error in the loop body")
+    end
+    a
 end
-@test _atthreads_static_schedule() == [1:nthreads();]
-@test_throws TaskFailedException @threads for i = 1:1; _atthreads_static_schedule(); end
+@test_throws "user error in the loop body" _atthreads_dynamic_with_error(zeros(nthreads()))
 
 try
     @macroexpand @threads(for i = 1:10, j = 1:10; end)