@c -*-texinfo-*-

@c This file is part of the StarPU Handbook.
@c Copyright (C) 2009--2011  Universit@'e de Bordeaux 1
@c Copyright (C) 2010, 2011, 2012, 2013  Centre National de la Recherche Scientifique
@c Copyright (C) 2011 Institut National de Recherche en Informatique et Automatique
@c See the file starpu.texi for copying conditions.

The integration of MPI transfers within task parallelism is done in a
very natural way by the means of asynchronous interactions between the
application and StarPU.  This is implemented in a separate libstarpumpi library
which basically provides "StarPU" equivalents of @code{MPI_*} functions, where
@code{void *} buffers are replaced with @code{starpu_data_handle_t}s, and all
GPU-RAM-NIC transfers are handled efficiently by StarPU-MPI.  The user has to
use the usual @code{mpirun} command of the MPI implementation to start StarPU on
the different MPI nodes.

An MPI Insert Task function provides an even more seamless transition to a
distributed application, by automatically issuing all required data transfers
according to the task graph and an application-provided distribution.

@menu
* Simple Example::
* Point to point communication::
* Exchanging User Defined Data Interface::
* MPI Insert Task Utility::
* MPI Collective Operations::
@end menu

@node Simple Example
@section Simple Example

The flags required to compile or link against the MPI layer are then
accessible with the following commands:

@example
$ pkg-config --cflags starpumpi-1.0  # options for the compiler
$ pkg-config --libs starpumpi-1.0    # options for the linker
@end example

Also pass the @code{--static} option if the application is to be linked statically.

@cartouche
@smallexample
void increment_token(void)
@{
    struct starpu_task *task = starpu_task_create();

    task->cl = &increment_cl;
    task->handles[0] = token_handle;

    starpu_task_submit(task);
@}
@end smallexample
@end cartouche

@cartouche
@smallexample
int main(int argc, char **argv)
@{
    int rank, size;

    starpu_init(NULL);
    starpu_mpi_initialize_extended(&rank, &size);

    starpu_vector_data_register(&token_handle, 0, (uintptr_t)&token, 1, sizeof(unsigned));

    unsigned nloops = NITER;
    unsigned loop;

    unsigned last_loop = nloops - 1;
    unsigned last_rank = size - 1;
@end smallexample
@end cartouche

@cartouche
@smallexample
    for (loop = 0; loop < nloops; loop++) @{
        int tag = loop*size + rank;

        if (loop == 0 && rank == 0)
        @{
            token = 0;
            fprintf(stdout, "Start with token value %d\n", token);
        @}
        else
        @{
            starpu_mpi_irecv_detached(token_handle, (rank+size-1)%size, tag,
                    MPI_COMM_WORLD, NULL, NULL);
        @}

        increment_token();

        if (loop == last_loop && rank == last_rank)
        @{
            starpu_data_acquire(token_handle, STARPU_R);
            fprintf(stdout, "Finished: token value %d\n", token);
            starpu_data_release(token_handle);
        @}
        else
        @{
            starpu_mpi_isend_detached(token_handle, (rank+1)%size, tag+1,
                    MPI_COMM_WORLD, NULL, NULL);
        @}
    @}

    starpu_task_wait_for_all();
@end smallexample
@end cartouche

@cartouche
@smallexample
    starpu_mpi_shutdown();
    starpu_shutdown();

    if (rank == last_rank)
    @{
        fprintf(stderr, "[%d] token = %d == %d * %d ?\n", rank, token, nloops, size);
        STARPU_ASSERT(token == nloops*size);
    @}
@end smallexample
@end cartouche

@node Point to point communication
@section Point to point communication

The standard point to point communications of MPI have been
implemented. The semantic is similar to the MPI one, but adapted to
the DSM provided by StarPU. A MPI request will only be submitted when
the data is available in the main memory of the node submitting the
request.

There is two types of asynchronous communications: the classic
asynchronous communications and the detached communications. The
classic asynchronous communications (@code{starpu_mpi_isend} and
@code{starpu_mpi_irecv}) need to be followed by a call to
@code{starpu_mpi_wait} or to @code{starpu_mpi_test} to wait for or to
test the completion of the communication. Waiting for or testing the
completion of detached communications is not possible, this is done
internally by StarPU-MPI, on completion, the resources are
automatically released. This mechanism is similar to the pthread
detach state attribute which determines whether a thread will be
created in a joinable or a detached state.

For any communication, the call of the function will result in the
creation of a StarPU-MPI request, the function
@code{starpu_data_acquire_cb} is then called to asynchronously request
StarPU to fetch the data in main memory; when the data is available in
main memory, a StarPU-MPI function is called to put the new request in
the list of the ready requests.

The StarPU-MPI progression thread regularly polls this list of ready
requests. For each new ready request, the appropriate function is
called to post the corresponding MPI call. For example, calling
@code{starpu_mpi_isend} will result in posting @code{MPI_Isend}. If
the request is marked as detached, the request will be put in the list
of detached requests.

The StarPU-MPI progression thread also polls the list of detached
requests. For each detached request, it regularly tests the completion
of the MPI request by calling @code{MPI_Test}. On completion, the data
handle is released, and if a callback was defined, it is called.

@ref{Communication} gives the list of all the point to point
communications defined in StarPU-MPI.

@node Exchanging User Defined Data Interface
@section Exchanging User Defined Data Interface

New data interfaces defined as explained in @ref{An example
of data interface} can also be used within StarPU-MPI and exchanged
between nodes. Two functions needs to be defined through
the type @code{struct starpu_data_interface_ops} (@pxref{Data
Interface API}). The pack function takes a handle and returns a
contiguous memory buffer along with its size where data to be conveyed to another node
should be copied. The reversed operation is implemented in the unpack
function which takes a contiguous memory buffer and recreates the data
handle.

@cartouche
@smallexample
static int complex_pack_data(starpu_data_handle_t handle, unsigned node, void **ptr, ssize_t *count)
@{
  STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));

  struct starpu_complex_interface *complex_interface =
    (struct starpu_complex_interface *) starpu_data_get_interface_on_node(handle, node);

  *count = complex_get_size(handle);
  *ptr = malloc(*count);
  memcpy(*ptr, complex_interface->real, complex_interface->nx*sizeof(double));
  memcpy(*ptr+complex_interface->nx*sizeof(double), complex_interface->imaginary,
         complex_interface->nx*sizeof(double));

  return 0;
@}
@end smallexample
@end cartouche

@cartouche
@smallexample
static int complex_unpack_data(starpu_data_handle_t handle, unsigned node, void *ptr, size_t count)
@{
  STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));

  struct starpu_complex_interface *complex_interface =
    (struct starpu_complex_interface *)	starpu_data_get_interface_on_node(handle, node);

  memcpy(complex_interface->real, ptr, complex_interface->nx*sizeof(double));
  memcpy(complex_interface->imaginary, ptr+complex_interface->nx*sizeof(double),
         complex_interface->nx*sizeof(double));

  return 0;
@}
@end smallexample
@end cartouche

@cartouche
@smallexample
static struct starpu_data_interface_ops interface_complex_ops =
@{
  ...
  .pack_data = complex_pack_data,
  .unpack_data = complex_unpack_data
@};
@end smallexample
@end cartouche

@page
@node MPI Insert Task Utility
@section MPI Insert Task Utility

To save the programmer from having to explicit all communications, StarPU
provides an "MPI Insert Task Utility". The principe is that the application
decides a distribution of the data over the MPI nodes by allocating it and
notifying StarPU of that decision, i.e. tell StarPU which MPI node "owns"
which data. It also decides, for each handle, an MPI tag which will be used to
exchange the content of the handle. All MPI nodes then process the whole task
graph, and StarPU automatically determines which node actually execute which
task, and trigger the required MPI transfers.

@deftypefun int starpu_data_set_tag (starpu_data_handle_t @var{handle}, int @var{tag})
Tell StarPU-MPI which MPI tag to use when exchanging the data.
@end deftypefun

@deftypefun int starpu_data_get_tag (starpu_data_handle_t @var{handle})
Returns the MPI tag to be used when exchanging the data.
@end deftypefun

@deftypefun int starpu_data_set_rank (starpu_data_handle_t @var{handle}, int @var{rank})
Tell StarPU-MPI which MPI node "owns" a given data, that is, the node which will
always keep an up-to-date value, and will by default execute tasks which write
to it.
@end deftypefun

@deftypefun int starpu_data_get_rank (starpu_data_handle_t @var{handle})
Returns the last value set by @code{starpu_data_set_rank}.
@end deftypefun

@defmac STARPU_EXECUTE_ON_NODE
this macro is used when calling @code{starpu_mpi_insert_task}, and
must be followed by a integer value which specified the node on which
to execute the codelet.
@end defmac

@defmac STARPU_EXECUTE_ON_DATA
this macro is used when calling @code{starpu_mpi_insert_task}, and
must be followed by a data handle to specify that the node owning the
given data will execute the codelet.
@end defmac

@deftypefun int starpu_mpi_insert_task (MPI_Comm @var{comm}, struct starpu_codelet *@var{codelet}, ...)
Create and submit a task corresponding to @var{codelet} with the following
arguments.  The argument list must be zero-terminated.

The arguments following the codelets are the same types as for the
function @code{starpu_insert_task} defined in @ref{Insert Task
Utility}. The extra argument @code{STARPU_EXECUTE_ON_NODE} followed by an
integer allows to specify the MPI node to execute the codelet. It is also
possible to specify that the node owning a specific data will execute
the codelet, by using @code{STARPU_EXECUTE_ON_DATA} followed by a data
handle.

The internal algorithm is as follows:
@enumerate
@item Find out which MPI node is going to execute the codelet.
      @enumerate
      @item If there is only one node owning data in W mode, it will
      be selected;
      @item If there is several nodes owning data in W node, the one
      selected will be the one having the least data in R mode so as
      to minimize the amount of data to be transfered;
      @item The argument @code{STARPU_EXECUTE_ON_NODE} followed by an
      integer can be used to specify the node;
      @item The argument @code{STARPU_EXECUTE_ON_DATA} followed by a
      data handle can be used to specify that the node owing the given
      data will execute the codelet.
      @end enumerate
@item Send and receive data as requested. Nodes owning data which need to be
read by the task are sending them to the MPI node which will execute it. The
latter receives them.
@item Execute the codelet. This is done by the MPI node selected in the
1st step of the algorithm.
@item If several MPI nodes own data to be written to, send written
data back to their owners.
@end enumerate

The algorithm also includes a communication cache mechanism that
allows not to send data twice to the same MPI node, unless the data
has been modified. The cache can be disabled
(@pxref{STARPU_MPI_CACHE}).
@c todo parler plus du cache

@end deftypefun

@deftypefun void starpu_mpi_get_data_on_node (MPI_Comm @var{comm}, starpu_data_handle_t @var{data_handle}, int @var{node})
Transfer data @var{data_handle} to MPI node @var{node}, sending it from its
owner if needed. At least the target node and the owner have to call the
function.
@end deftypefun

Here an stencil example showing how to use @code{starpu_mpi_insert_task}. One
first needs to define a distribution function which specifies the
locality of the data. Note that that distribution information needs to
be given to StarPU by calling @code{starpu_data_set_rank}.

@cartouche
@smallexample
/* Returns the MPI node number where data is */
int my_distrib(int x, int y, int nb_nodes) @{
  /* Block distrib */
  return ((int)(x / sqrt(nb_nodes) + (y / sqrt(nb_nodes)) * sqrt(nb_nodes))) % nb_nodes;

  // /* Other examples useful for other kinds of computations */
  // /* / distrib */
  // return (x+y) % nb_nodes;

  // /* Block cyclic distrib */
  // unsigned side = sqrt(nb_nodes);
  // return x % side + (y % side) * size;
@}
@end smallexample
@end cartouche

Now the data can be registered within StarPU. Data which are not
owned but will be needed for computations can be registered through
the lazy allocation mechanism, i.e. with a @code{home_node} set to -1.
StarPU will automatically allocate the memory when it is used for the
first time.

One can note an optimization here (the @code{else if} test): we only register
data which will be needed by the tasks that we will execute.

@cartouche
@smallexample
    unsigned matrix[X][Y];
    starpu_data_handle_t data_handles[X][Y];

    for(x = 0; x < X; x++) @{
        for (y = 0; y < Y; y++) @{
            int mpi_rank = my_distrib(x, y, size);
             if (mpi_rank == my_rank)
                /* Owning data */
                starpu_variable_data_register(&data_handles[x][y], 0,
                                              (uintptr_t)&(matrix[x][y]), sizeof(unsigned));
            else if (my_rank == my_distrib(x+1, y, size) || my_rank == my_distrib(x-1, y, size)
                  || my_rank == my_distrib(x, y+1, size) || my_rank == my_distrib(x, y-1, size))
                /* I don't own that index, but will need it for my computations */
                starpu_variable_data_register(&data_handles[x][y], -1,
                                              (uintptr_t)NULL, sizeof(unsigned));
            else
                /* I know it's useless to allocate anything for this */
                data_handles[x][y] = NULL;
            if (data_handles[x][y])
                starpu_data_set_rank(data_handles[x][y], mpi_rank);
        @}
    @}
@end smallexample
@end cartouche

Now @code{starpu_mpi_insert_task()} can be called for the different
steps of the application.

@cartouche
@smallexample
    for(loop=0 ; loop<niter; loop++)
        for (x = 1; x < X-1; x++)
            for (y = 1; y < Y-1; y++)
                starpu_mpi_insert_task(MPI_COMM_WORLD, &stencil5_cl,
                                       STARPU_RW, data_handles[x][y],
                                       STARPU_R, data_handles[x-1][y],
                                       STARPU_R, data_handles[x+1][y],
                                       STARPU_R, data_handles[x][y-1],
                                       STARPU_R, data_handles[x][y+1],
                                       0);
    starpu_task_wait_for_all();
@end smallexample
@end cartouche

I.e. all MPI nodes process the whole task graph, but as mentioned above, for
each task, only the MPI node which owns the data being written to (here,
@code{data_handles[x][y]}) will actually run the task. The other MPI nodes will
automatically send the required data.

This can be a concern with a growing number of nodes. To avoid this, the
application can prune the task for loops according to the data distribution,
so as to only submit tasks on nodes which have to care about them (either to
execute them, or to send the required data).

@node MPI Collective Operations
@section MPI Collective Operations

@deftypefun int starpu_mpi_scatter_detached (starpu_data_handle_t *@var{data_handles}, int @var{count}, int @var{root}, MPI_Comm @var{comm}, {void (*}@var{scallback})(void *), {void *}@var{sarg}, {void (*}@var{rcallback})(void *), {void *}@var{rarg})
Scatter data among processes of the communicator based on the ownership of
the data. For each data of the array @var{data_handles}, the
process @var{root} sends the data to the process owning this data.
Processes receiving data must have valid data handles to receive them.
On completion of the collective communication, the @var{scallback} function is
called with the argument @var{sarg} on the process @var{root}, the @var{rcallback} function is
called with the argument @var{rarg} on any other process.
@end deftypefun

@deftypefun int starpu_mpi_gather_detached (starpu_data_handle_t *@var{data_handles}, int @var{count}, int @var{root}, MPI_Comm @var{comm}, {void (*}@var{scallback})(void *), {void *}@var{sarg}, {void (*}@var{rcallback})(void *), {void *}@var{rarg})
Gather data from the different processes of the communicator onto the
process @var{root}. Each process owning data handle in the array
@var{data_handles} will send them to the process @var{root}. The
process @var{root} must have valid data handles to receive the data.
On completion of the collective communication, the @var{rcallback} function is
called with the argument @var{rarg} on the process @var{root}, the @var{scallback} function is
called with the argument @var{sarg} on any other process.
@end deftypefun

@page
@cartouche
@smallexample
if (rank == root)
@{
    /* Allocate the vector */
    vector = malloc(nblocks * sizeof(float *));
    for(x=0 ; x<nblocks ; x++)
    @{
        starpu_malloc((void **)&vector[x], block_size*sizeof(float));
    @}
@}

/* Allocate data handles and register data to StarPU */
data_handles = malloc(nblocks*sizeof(starpu_data_handle_t *));
for(x = 0; x < nblocks ;  x++)
@{
    int mpi_rank = my_distrib(x, nodes);
    if (rank == root) @{
        starpu_vector_data_register(&data_handles[x], 0, (uintptr_t)vector[x],
                                    blocks_size, sizeof(float));
    @}
    else if ((mpi_rank == rank) || ((rank == mpi_rank+1 || rank == mpi_rank-1))) @{
        /* I own that index, or i will need it for my computations */
        starpu_vector_data_register(&data_handles[x], -1, (uintptr_t)NULL,
                                   block_size, sizeof(float));
    @}
    else @{
        /* I know it's useless to allocate anything for this */
        data_handles[x] = NULL;
    @}
    if (data_handles[x]) @{
        starpu_data_set_rank(data_handles[x], mpi_rank);
    @}
@}

/* Scatter the matrix among the nodes */
starpu_mpi_scatter_detached(data_handles, nblocks, root, MPI_COMM_WORLD);

/* Calculation */
for(x = 0; x < nblocks ;  x++) @{
    if (data_handles[x]) @{
        int owner = starpu_data_get_rank(data_handles[x]);
        if (owner == rank) @{
            starpu_insert_task(&cl, STARPU_RW, data_handles[x], 0);
        @}
    @}
@}

/* Gather the matrix on main node */
starpu_mpi_gather_detached(data_handles, nblocks, 0, MPI_COMM_WORLD);
@end smallexample
@end cartouche