Multiple Completions

Multiple Completions


Up: Nonblocking communication Next: Probe and Cancel Previous: Semantics of Nonblocking Communications

It is convenient to be able to wait for the completion of any, some, or all the operations in a list, rather than having to wait for a specific message. A call to MPI_WAITANY or MPI_TESTANY can be used to wait for the completion of one out of several operations. A call to MPI_WAITALL or MPI_TESTALL can be used to wait for all pending operations in a list. A call to MPI_WAITSOME or MPI_TESTSOME can be used to complete all enabled operations in a list.

MPI_WAITANY (count, array_of_requests, index, status)
[ IN count] list length (integer)
[ INOUT array_of_requests] array of requests (array of handles)
[ OUT index] index of handle for operation that completed (integer)
[ OUT status] status object (Status)

int MPI_Waitany(int count, MPI_Request *array_of_requests, int *index, MPI_Status *status)

MPI_WAITANY(COUNT, ARRAY_OF_REQUESTS, INDEX, STATUS, IERROR)
INTEGER COUNT, ARRAY_OF_REQUESTS(*), INDEX, STATUS(MPI_STATUS_SIZE), IERROR

Blocks until one of the operations associated with the active requests in the array has completed. If more then one operation is enabled and can terminate, one is arbitrarily chosen. Returns in index the index of that request in the array and returns in status the status of the completing communication. (The array is indexed from zero in C, and from one in Fortran.) If the request was allocated by a nonblocking communication operation, then it is deallocated and the request handle is set to MPI_REQUEST_NULL.

The array_of_requests list may contain null or inactive handles. If the list contains no active handles (list has length zero or all entries are null or inactive), then the call returns immediately with index = MPI_UNDEFINED, and a empty status.

The execution of MPI_WAITANY(count, array_of_requests, index, status) has the same effect as the execution of MPI_WAIT(&array_of_requests[i], status), where i is the value returned by index (unless the value of index is MPI_UNDEFINED). MPI_WAITANY with an array containing one active entry is equivalent to MPI_WAIT.

MPI_TESTANY(count, array_of_requests, index, flag, status)
[ IN count] list length (integer)
[ INOUT array_of_requests] array of requests (array of handles)
[ OUT index] index of operation that completed, or MPI_UNDEFINED if none completed (integer)
[ OUT flag] true if one of the operations is complete (logical)
[ OUT status] status object (Status)

int MPI_Testany(int count, MPI_Request *array_of_requests, int *index, int *flag, MPI_Status *status)

MPI_TESTANY(COUNT, ARRAY_OF_REQUESTS, INDEX, FLAG, STATUS, IERROR)
LOGICAL FLAG
INTEGER COUNT, ARRAY_OF_REQUESTS(*), INDEX, STATUS(MPI_STATUS_SIZE), IERROR

Tests for completion of either one or none of the operations associated with active handles. In the former case, it returns flag = true, returns in index the index of this request in the array, and returns in status the status of that operation; if the request was allocated by a nonblocking communication call then the request is deallocated and the handle is set to MPI_REQUEST_NULL. (The array is indexed from zero in C, and from one in Fortran.) In the latter case (no operation completed), it returns flag = false, returns a value of MPI_UNDEFINED in index and status is undefined.

The array may contain null or inactive handles. If the array contains no active handles then the call returns immediately with flag = true, index = MPI_UNDEFINED, and an empty status.

If the array of requests contains active handles then the execution of MPI_TESTANY(count, array_of_requests, index, status) has the same effect as the execution of MPI_TEST( &array_of_requests[i], flag, status), for i=0, 1 ,..., count-1, in some arbitrary order, until one call returns flag = true, or all fail. In the former case, index is set to the last value of i, and in the latter case, it is set to MPI_UNDEFINED. MPI_TESTANY with an array containing one active entry is equivalent to MPI_TEST.


[] Rationale.

The function MPI_TESTANY returns with flag = true exactly in those situations where the function MPI_WAITANY returns; both functions return in that case the same values in the remaining parameters. Thus, a blocking MPI_WAITANY can be easily replaced by a nonblocking MPI_TESTANY. The same relation holds for the other pairs of Wait and Test functions defined in this section. ( End of rationale.)

MPI_WAITALL( count, array_of_requests, array_of_statuses)
[ IN count] lists length (integer)
[ INOUT array_of_requests] array of requests (array of handles)
[ OUT array_of_statuses] array of status objects (array of Status)

int MPI_Waitall(int count, MPI_Request *array_of_requests, MPI_Status *array_of_statuses)

MPI_WAITALL(COUNT, ARRAY_OF_REQUESTS, ARRAY_OF_STATUSES, IERROR)
INTEGER COUNT, ARRAY_OF_REQUESTS(*)
INTEGER ARRAY_OF_STATUSES(MPI_STATUS_SIZE,*), IERROR

Blocks until all communication operations associated with active handles in the list complete, and return the status of all these operations (this includes the case where no handle in the list is active). Both arrays have the same number of valid entries. The i-th entry in array_of_statuses is set to the return status of the i-th operation. Requests that were created by nonblocking communication operations are deallocated and the corresponding handles in the array are set to MPI_REQUEST_NULL. The list may contain null or inactive handles. The call sets to empty the status of each such entry.

The error-free execution of MPI_WAITALL(count, array_of_requests, array_of_statuses) has the same effect as the execution of
MPI_WAIT(&array_of_request[i], &array_of_statuses[i]), for i=0 ,..., count-1, in some arbitrary order. MPI_WAITALL with an array of length one is equivalent to MPI_WAIT.

When one or more of the communications completed by a call to MPI_WAITALL fail, it is desireable to return specific information on each communication. The function MPI_WAITALL will return in such case the error code MPI_ERR_IN_STATUS and will set the error field of each status to a specific error code. This code will be MPI_SUCCESS, if the specific communication completed; it will be another specific error code, if it failed; or it can be MPI_ERR_PENDING if it has neither failed nor completed. The function MPI_WAITALL will return MPI_SUCCESS if no request had an error, or will return another error code if it failed for other reasons (such as invalid arguments). In such cases, it will not update the error fields of the statuses.


[] Rationale.

This design streamlines error handling in the application. The application code need only test the (single) function result to determine if an error has occurred. It needs to check each individual status only when an error occurred. ( End of rationale.)

MPI_TESTALL(count, array_of_requests, flag, array_of_statuses)
[ IN count] lists length (integer)
[ INOUT array_of_requests] array of requests (array of handles)
[ OUT flag] (logical)
[ OUT array_of_statuses] array of status objects (array of Status)

int MPI_Testall(int count, MPI_Request *array_of_requests, int *flag, MPI_Status *array_of_statuses)

MPI_TESTALL(COUNT, ARRAY_OF_REQUESTS, FLAG, ARRAY_OF_STATUSES, IERROR)
LOGICAL FLAG
INTEGER COUNT, ARRAY_OF_REQUESTS(*), ARRAY_OF_STATUSES(MPI_STATUS_SIZE,*), IERROR

Returns flag = true if all communications associated with active handles in the array have completed (this includes the case where no handle in the list is active). In this case, each status entry that corresponds to an active handle request is set to the status of the corresponding communication; if the request was allocated by a nonblocking communication call then it is deallocated, and the handle is set to MPI_REQUEST_NULL. Each status entry that corresponds to a null or inactive handle is set to empty.

Otherwise, flag = false is returned, no request is modified and the values of the status entries are undefined. This is a local operation.

Errors that occurred during the execution of MPI_TESTALL are handled as errors in MPI_WAITALL.

MPI_WAITSOME(incount, array_of_requests, outcount, array_of_indices, array_of_statuses)
[ IN incount] length of array_of_requests (integer)
[ INOUT array_of_requests] array of requests (array of handles)
[ OUT outcount] number of completed requests (integer)
[ OUT array_of_indices] array of indices of operations that completed (array of integers)
[ OUT array_of_statuses] array of status objects for operations that completed (array of Status)

int MPI_Waitsome(int incount, MPI_Request *array_of_requests, int *outcount, int *array_of_indices, MPI_Status *array_of_statuses)

MPI_WAITSOME(INCOUNT, ARRAY_OF_REQUESTS, OUTCOUNT, ARRAY_OF_INDICES, ARRAY_OF_STATUSES, IERROR)
INTEGER INCOUNT, ARRAY_OF_REQUESTS(*), OUTCOUNT, ARRAY_OF_INDICES(*), ARRAY_OF_STATUSES(MPI_STATUS_SIZE,*), IERROR

Waits until at least one of the operations associated with active handles in the list have completed. Returns in outcount the number of requests from the list array_of_requests that have completed. Returns in the first outcount locations of the array array_of_indices the indices of these operations (index within the array array_of_requests; the array is indexed from zero in C and from one in Fortran). Returns in the first outcount locations of the array array_of_status the status for these completed operations. If a request that completed was allocated by a nonblocking communication call, then it is deallocated, and the associated handle is set to MPI_REQUEST_NULL.

If the list contains no active handles, then the call returns immediately with outcount = MPI_UNDEFINED.

When one or more of the communications completed by MPI_WAITSOME fails, then it is desirable to return specific information on each communication. The arguments outcount, array_of_indices and array_of_statuses will be adjusted to indicate completion of all communications that have succeeded or failed. The call will return the error code MPI_ERR_IN_STATUS and the error field of each status returned will be set to indicate success or to indicate the specific error that occurred. The call will return MPI_SUCCESS if no request resulted in an error, and will return another error code if it failed for other reasons (such as invalid arguments). In such cases, it will not update the error fields of the statuses.

MPI_TESTSOME(incount, array_of_requests, outcount, array_of_indices, array_of_statuses)
[ IN incount] length of array_of_requests (integer)
[ INOUT array_of_requests] array of requests (array of handles)
[ OUT outcount] number of completed requests (integer)
[ OUT array_of_indices] array of indices of operations that completed (array of integers)
[ OUT array_of_statuses] array of status objects for operations that completed (array of Status)

int MPI_Testsome(int incount, MPI_Request *array_of_requests, int *outcount, int *array_of_indices, MPI_Status *array_of_statuses)

MPI_TESTSOME(INCOUNT, ARRAY_OF_REQUESTS, OUTCOUNT, ARRAY_OF_INDICES, ARRAY_OF_STATUSES, IERROR)
INTEGER INCOUNT, ARRAY_OF_REQUESTS(*), OUTCOUNT, ARRAY_OF_INDICES(*), ARRAY_OF_STATUSES(MPI_STATUS_SIZE,*), IERROR

Behaves like MPI_WAITSOME, except that it returns immediately. If no operation has completed it returns outcount = 0. If there is no active handle in the list it returns outcount = MPI_UNDEFINED.

MPI_TESTSOME is a local operation, which returns immediately, whereas MPI_WAITSOME will block until a communication completes, if it was passed a list that contains at least one active handle. Both calls fulfill a fairness requirement: If a request for a receive repeatedly appears in a list of requests passed to MPI_WAITSOME or MPI_TESTSOME, and a matching send has been posted, then the receive will eventually succeed, unless the send is satisfied by another receive; and similarly for send requests.

Errors that occur during the execution of MPI_TESTSOME are handled as for
MPI_WAITSOME.


[] Advice to users.

The use of MPI_TESTSOME is likely to be more efficient than the use of MPI_TESTANY. The former returns information on all completed communications, with the latter, a new call is required for each communication that completes.

A server with multiple clients can use MPI_WAITSOME so as not to starve any client. Clients send messages to the server with service requests. The server calls MPI_WAITSOME with one receive request for each client, and then handles all receives that completed. If a call to MPI_WAITANY is used instead, then one client could starve while requests from another client always sneak in first. ( End of advice to users.)

[] Advice to implementors.

MPI_TESTSOME should complete as many pending communications as possible. ( End of advice to implementors.)

  
 
Client-server code (starvation can occur). 
 
CALL MPI_COMM_SIZE(comm, size, ierr) 
CALL MPI_COMM_RANK(comm, rank, ierr) 
IF(rank > 0) THEN         ! client code 
    DO WHILE(.TRUE.) 
       CALL MPI_ISEND(a, n, MPI_REAL, 0, tag, comm, request, ierr) 
       CALL MPI_WAIT(request, status, ierr) 
    END DO 
ELSE         ! rank=0 -- server code 
       DO i=1, size-1 
          CALL MPI_IRECV(a(1,i), n, MPI_REAL, i tag, 
                   comm, request_list(i), ierr) 
       END DO 
       DO WHILE(.TRUE.) 
          CALL MPI_WAITANY(size-1, request_list, index, status, ierr) 
          CALL DO_SERVICE(a(1,index))  ! handle one message 
          CALL MPI_IRECV(a(1, index), n, MPI_REAL, index, tag, 
                    comm, request_list(index), ierr) 
       END DO 
END IF 


  
 
Same code, using  MPI_WAITSOME. 
 
CALL MPI_COMM_SIZE(comm, size, ierr) 
CALL MPI_COMM_RANK(comm, rank, ierr) 
IF(rank > 0) THEN         ! client code 
    DO WHILE(.TRUE.) 
       CALL MPI_ISEND(a, n, MPI_REAL, 0, tag, comm, request, ierr) 
       CALL MPI_WAIT(request, status, ierr) 
    END DO 
ELSE         ! rank=0 -- server code 
    DO i=1, size-1 
       CALL MPI_IRECV(a(1,i), n, MPI_REAL, i, tag, 
                      comm, requests(i), ierr) 
    END DO 
    DO WHILE(.TRUE.) 
       CALL MPI_WAITSOME(size, request_list, numdone, 
                        indices, statuses, ierr) 
       DO i=1, numdone 
          CALL DO_SERVICE(a(1, indices(i))) 
          CALL MPI_IRECV(a(1, indices(i)), n, MPI_REAL, 0, tag, 
                       comm, requests(indices(i)), ierr) 
       END DO 
    END DO 
END IF 



Up: Nonblocking communication Next: Probe and Cancel Previous: Semantics of Nonblocking Communications


Return to MPI Standard Index
Return to MPI home page