updates to docs and added a new option to not broadcast results

README.md

@@ -6,8 +6,8 @@
 [codecov-img]: https://codecov.io/gh/psrenergy/JobQueueMPI.jl/coverage.svg?branch=master
 [codecov-url]: https://codecov.io/gh/psrenergy/JobQueueMPI.jl?branch=master
 
-| **Build Status** | **Coverage** | 
-|:-----------------:|:-----------------:|
+| **Build Status** | **Coverage** | **Documentation** |
+|:-----------------:|:-----------------:|:-----------------:|
 | [![Build Status][build-img]][build-url] | [![Codecov branch][codecov-img]][codecov-url] |[![](https://img.shields.io/badge/docs-latest-blue.svg)](https://psrenergy.github.io/JobQueueMPI.jl/dev/)
 
 

docs/make.jl

@@ -12,10 +12,10 @@ makedocs(;
     authors = "psrenergy",
     pages = [
         "Home" => "index.md",
-    ]
+    ],
 )
 
 deploydocs(;
     repo = "github.com/psrenergy/JobQueueMPI.jl.git",
     push_preview = true,
-)
+)

docs/src/index.md

@@ -19,6 +19,100 @@ JobQueueMPI.jl has the following components:
 - `Controller`: The controller is responsible for managing the jobs and the workers. It keeps track of the jobs that have been sent and received and sends the jobs to the available workers.
 - `Worker`: The worker is responsible for executing the jobs. It receives the jobs from the controller, executes them, and sends the results back to the controller.
 
+Users can call functions to compute jobs in parallel in two ways:
+ - Building a function and using a `pmap` implementation that will put the function in the job queue and send it to the workers.
+```julia
+using JobQueueMPI
+
+function sum_100(value)
+    return value + 100
+end
+
+sum_100_answer = JobQueueMPI.pmap(sum_100, collect(1:10))
+```
+ - Building the jobs and sending them to workers explicitly. There are examples of this structure in the test folder. This way is much more flexible than the first one, but it requires more code and knowledge about how MPI works.
+
+```julia
+using JobQueueMPI
+
+mutable struct Message
+    value::Int
+    vector_idx::Int
+end
+
+all_jobs_done(controller) = JQM.is_job_queue_empty(controller) && !JQM.any_pending_jobs(controller)
+
+function sum_100(message::Message)
+    message.value += 100
+    return message
+end
+
+function update_data(new_data, message::Message)
+    idx = message.vector_idx
+    value = message.value
+    return new_data[idx] = value
+end
+
+function workers_loop()
+    if JQM.is_worker_process()
+        worker = JQM.Worker()
+        while true
+            job = JQM.receive_job(worker)
+            message = JQM.get_message(job)
+            if message == JQM.TerminationMessage()
+                break
+            end
+            return_message = sum_100(message)
+            JQM.send_job_answer_to_controller(worker, return_message)
+        end
+        exit(0)
+    end
+end
+
+function job_queue(data)
+    JQM.mpi_init()
+    JQM.mpi_barrier()
+
+    T = eltype(data)
+    N = length(data)
+
+    if JQM.is_controller_process()
+        new_data = Array{T}(undef, N)
+
+        controller = JQM.Controller(JQM.num_workers())
+
+        for i in eachindex(data)
+            message = Message(data[i], i)
+            JQM.add_job_to_queue!(controller, message)
+        end
+
+        while !all_jobs_done(controller)
+            if !JQM.is_job_queue_empty(controller)
+                JQM.send_jobs_to_any_available_workers(controller)
+            end
+            if JQM.any_pending_jobs(controller)
+                job_answer = JQM.check_for_job_answers(controller)
+                if !isnothing(job_answer)
+                    message = JQM.get_message(job_answer)
+                    update_data(new_data, message)
+                end
+            end
+        end
+
+        JQM.send_termination_message()
+
+        return new_data
+    end
+    workers_loop()
+    JQM.mpi_barrier()
+    JQM.mpi_finalize()
+    return nothing
+end
+
+data = collect(1:10)
+new_data = job_queue(data)
+```
+
 ## API
 
 ```@docs

src/pmap.jl

@@ -39,10 +39,10 @@ The controller process will return the answer in the same order as the jobs were
 return nothing.
 """
 function pmap(
-    f::Function, 
-    jobs::Vector, 
-    data_defined_in_process = nothing; 
-    return_result_in_all_processes::Bool = false
+    f::Function,
+    jobs::Vector,
+    data_defined_in_process = nothing;
+    return_result_in_all_processes::Bool = true,
 )
     result = Vector{Any}(undef, length(jobs))
     if is_running_in_parallel()

test/test_pmap_mpi.jl

@@ -24,7 +24,7 @@ divisors_answer = JQM.pmap(get_divisors, collect(1:10))
 @testset "pmap MPI" begin
     @test sum_100_answer == [101, 102, 103, 104, 105, 106, 107, 108, 109, 110]
     @test divisors_answer ==
-            [[1], [1, 2], [1, 3], [1, 2, 4], [1, 5], [1, 2, 3, 6], [1, 7], [1, 2, 4, 8], [1, 3, 9], [1, 2, 5, 10]]
+          [[1], [1, 2], [1, 3], [1, 2, 4], [1, 5], [1, 2, 3, 6], [1, 7], [1, 2, 4, 8], [1, 3, 9], [1, 2, 5, 10]]
 end
 
 JQM.mpi_finalize()

test/test_pmap_mpi_optim.jl

@@ -8,19 +8,18 @@ JQM.mpi_init()
 N = 5
 data = collect(1:N)
 function g(x, i, data)
-    return i * (x[i] - 2 * i) ^ 2 + data[i]
+    return i * (x[i] - 2 * i)^2 + data[i]
 end
 x0 = zeros(N)
 list_i = collect(1:N)
 fake_input = Int[] # ignored
 
-
 let
     is_done = false
     if JQM.is_controller_process()
         ret = optimize(x0, NelderMead()) do x
             MPI.bcast(is_done, MPI.COMM_WORLD)
-            g_x = JQM.pmap((v)->g(v[1], v[2], data), [(x, i) for i in list_i])
+            g_x = JQM.pmap((v) -> g(v[1], v[2], data), [(x, i) for i in list_i])
             return sum(g_x)
         end
         # tell workers to stop calling pmap
@@ -42,11 +41,11 @@ let
             if is_done
                 break
             end
-            JQM.pmap((v)->g(v[1], v[2], data), fake_input)
+            JQM.pmap((v) -> g(v[1], v[2], data), fake_input)
         end
     end
 end
 
 JQM.mpi_barrier()
 
-JQM.mpi_finalize()
+JQM.mpi_finalize()