Merge pull request #373 from JuliaParallel/jps/task-balance

Implement work stealing

Author: jpsamaroo

Project.toml

@@ -5,6 +5,7 @@ version = "0.16.3"
 [deps]
 Colors = "5ae59095-9a9b-59fe-a467-6f913c188581"
 ContextVariablesX = "6add18c4-b38d-439d-96f6-d6bc489c04c5"
+DataStructures = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8"
 Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"

docs/src/index.md

@@ -227,5 +227,41 @@ opts = Dagger.Sch.ThunkOptions(;single=1)
 delayed(+)(1, 2; options=opts)
 ```
 
+### Core vs. Worker Schedulers
+
+Dagger's scheduler is really two kinds of entities: the "core" scheduler, and
+"worker" schedulers:
+
+The core scheduler runs on worker 1, thread 1, and is the entrypoint to tasks
+which have been submitted. The core scheduler manages all task dependencies,
+notifies calls to `wait` and `fetch` of task completion, and generally performs
+initial task placement. The core scheduler has cached information about each
+worker and their processors, and uses that information (together with metrics
+about previous tasks and other aspects of the Dagger runtime) to generate a
+near-optimal just-in-time task schedule.
+
+The worker schedulers each run as a set of tasks across all workers and all
+processors, and handles data movement and task execution. Once the core
+scheduler has scheduled and launched a task, it arrives at the worker scheduler
+for handling. The worker scheduler will pass the task to a queue for the
+assigned processor, where it will wait until the processor has a sufficient
+amount of "occupancy" for the task. Once the processor is ready for the task,
+it will first fetch all arguments to the task from other workers, and then it
+will execute the task, package the result into a `Chunk`, and pass that back to
+the core scheduler.
+
+### Workload Balancing
+
+In general, Dagger's core scheduler tries to balance workloads as much as
+possible across all the available processors, but it can fail to do so
+effectively when either the cached per-processor information is outdated, or
+when the estimates about the task's behavior are inaccurate. To minimize the
+impact of this potential workload imbalance, the worker schedulers' processors
+will attempt to steal tasks from each other when they are under-occupied. Tasks
+will only be stolen if their [scope](`Scopes`) matches the processor attempting
+the steal, so tasks with wider scopes have better balancing potential.
+
+### Scheduler/Thunk Options
+
 [`Dagger.Sch.SchedulerOptions`](@ref)
 [`Dagger.Sch.ThunkOptions`](@ref)

src/Dagger.jl

@@ -36,6 +36,7 @@ include("chunks.jl")
 include("compute.jl")
 include("utils/clock.jl")
 include("utils/system_uuid.jl")
+include("utils/locked-object.jl")
 include("sch/Sch.jl"); using .Sch
 
 # Array computations

src/processor.jl

@@ -1,5 +1,7 @@
 export OSProc, Context, addprocs!, rmprocs!
 
+import Base: @invokelatest
+
 """
     Processor
 
@@ -157,8 +159,9 @@ function execute!(proc::ThreadProc, @nospecialize(f), @nospecialize(args...))
     task = Task() do
         set_tls!(tls)
         TimespanLogging.prof_task_put!(tls.sch_handle.thunk_id.id)
-        f(args...)
+        @invokelatest f(args...)
     end
+    task.sticky = true
     ret = ccall(:jl_set_task_tid, Cint, (Any, Cint), task, proc.tid-1)
     if ret == 0
         error("jl_set_task_tid == 0")
@@ -310,8 +313,7 @@ get_tls() = (
     sch_uid=task_local_storage(:_dagger_sch_uid),
     sch_handle=task_local_storage(:_dagger_sch_handle),
     processor=thunk_processor(),
-    time_utilization=task_local_storage(:_dagger_time_utilization),
-    alloc_utilization=task_local_storage(:_dagger_alloc_utilization),
+    task_spec=task_local_storage(:_dagger_task_spec),
 )
 
 """
@@ -323,6 +325,5 @@ function set_tls!(tls)
     task_local_storage(:_dagger_sch_uid, tls.sch_uid)
     task_local_storage(:_dagger_sch_handle, tls.sch_handle)
     task_local_storage(:_dagger_processor, tls.processor)
-    task_local_storage(:_dagger_time_utilization, tls.time_utilization)
-    task_local_storage(:_dagger_alloc_utilization, tls.alloc_utilization)
+    task_local_storage(:_dagger_task_spec, tls.task_spec)
 end