The examples in this section illustrate the application of the MPI consistency and semantics guarantees. These address
/* Process 0 */ int i, a[10] ; int TRUE = 1;for ( i=0;i<10;i++) a[i] = 5 ;
MPI_File_open( MPI_COMM_WORLD, "workfile", MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh0 ) ; MPI_File_set_view( fh0, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; MPI_File_set_atomicity( fh0, TRUE ) ; MPI_File_write_at(fh0, 0, a, 10, MPI_INT, &status) ; /* MPI_Barrier( MPI_COMM_WORLD ) ; */
/* Process 1 */ int b[10] ; int TRUE = 1; MPI_File_open( MPI_COMM_WORLD, "workfile", MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh1 ) ; MPI_File_set_view( fh1, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; MPI_File_set_atomicity( fh1, TRUE ) ; /* MPI_Barrier( MPI_COMM_WORLD ) ; */ MPI_File_read_at(fh1, 0, b, 10, MPI_INT, &status) ;A user may guarantee that the write on process 0 precedes the read on process 1 by imposing temporal order with, for example, calls to MPI_BARRIER.
[] Advice to users.
Routines other than MPI_BARRIER may be used to impose
temporal order. In the example above, process 0 could use MPI_SEND
to send a 0 byte message, received by process 1 using MPI_RECV.
( End of advice to users.)
Alternatively, a user can impose consistency with nonatomic mode set:
/* Process 0 */ int i, a[10] ; for ( i=0;i<10;i++) a[i] = 5 ;MPI_File_open( MPI_COMM_WORLD, "workfile", MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh0 ) ; MPI_File_set_view( fh0, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; MPI_File_write_at(fh0, 0, a, 10, MPI_INT, &status ) ; MPI_File_sync( fh0 ) ; MPI_Barrier( MPI_COMM_WORLD ) ; MPI_File_sync( fh0 ) ;
/* Process 1 */ int b[10] ; MPI_File_open( MPI_COMM_WORLD, "workfile", MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh1 ) ; MPI_File_set_view( fh1, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; MPI_File_sync( fh1 ) ; MPI_Barrier( MPI_COMM_WORLD ) ; MPI_File_sync( fh1 ) ; MPI_File_read_at(fh1, 0, b, 10, MPI_INT, &status ) ;The ``sync-barrier-sync'' construct is required because:
/* ---------------- THIS EXAMPLE IS ERRONEOUS --------------- */ /* Process 0 */ int i, a[10] ; for ( i=0;i<10;i++) a[i] = 5 ;MPI_File_open( MPI_COMM_WORLD, "workfile", MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh0 ) ; MPI_File_set_view( fh0, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; MPI_File_write_at(fh0, 0, a, 10, MPI_INT, &status ) ; MPI_File_sync( fh0 ) ; MPI_Barrier( MPI_COMM_WORLD ) ;
/* Process 1 */ int b[10] ; MPI_File_open( MPI_COMM_WORLD, "workfile", MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh1 ) ; MPI_File_set_view( fh1, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; MPI_Barrier( MPI_COMM_WORLD ) ; MPI_File_sync( fh1 ) ; MPI_File_read_at(fh1, 0, b, 10, MPI_INT, &status ) ;The above program also violates the MPI rule against out-of-order collective operations and will deadlock for implementations in which MPI_FILE_SYNC blocks./* ---------------- THIS EXAMPLE IS ERRONEOUS --------------- */
[] Advice to users.
Some implementations may choose to implement MPI_FILE_SYNC
as a temporally synchronizing function. When using such an
implementation, the ``sync-barrier-sync'' construct above can
be replaced by a single ``sync.'' The results of using such
code with an implementation for which MPI_FILE_SYNC is not
temporally synchronizing is undefined.
( End of advice to users.)
The behavior of asynchronous I/O operations is determined by applying the rules specified above for synchronous I/O operations.
The following examples all access a preexisting file ``myfile.'' Word 10 in myfile initially contains the integer 2. Each example writes and reads word 10.
First consider the following code fragment:
int a = 4, b, TRUE=1; MPI_File_open( MPI_COMM_WORLD, "myfile", MPI_MODE_RDWR, MPI_INFO_NULL, &fh ) ; MPI_File_set_view( fh, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; /* MPI_File_set_atomicity( fh, TRUE ) ; Use this to set atomic mode. */ MPI_File_iwrite_at(fh, 10, &a, 1, MPI_INT, &reqs[0]) ; MPI_File_iread_at(fh, 10, &b, 1, MPI_INT, &reqs[1]) ; MPI_Waitall(2, reqs, statuses) ;For asynchronous data access operations, MPI specifies that the access occurs at any time between the call to the asynchronous data access routine and the return from the corresponding request complete routine. Thus, executing either the read before the write, or the write before the read is consistent with program order. If atomic mode is set, then MPI guarantees sequential consistency, and the program will read either 2 or 4 into b. If atomic mode is not set, then sequential consistency is not guaranteed and the program may read something other than 2 or 4 due to the conflicting data access.
Similarly, the following code fragment does not order file accesses:
int a = 4, b; MPI_File_open( MPI_COMM_WORLD, "myfile", MPI_MODE_RDWR, MPI_INFO_NULL, &fh ) ; MPI_File_set_view( fh, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; /* MPI_File_set_atomicity( fh, TRUE ) ; Use this to set atomic mode. */ MPI_File_iwrite_at(fh, 10, &a, 1, MPI_INT, &reqs[0]) ; MPI_File_iread_at(fh, 10, &b, 1, MPI_INT, &reqs[1]) ; MPI_Wait(&reqs[0], &status) ; MPI_Wait(&reqs[1], &status) ;If atomic mode is set, either 2 or 4 will be read into b. Again, MPI does not guarantee sequential consistency in nonatomic mode.
On the other hand, the following code fragment:
int a = 4, b; MPI_File_open( MPI_COMM_WORLD, "myfile", MPI_MODE_RDWR, MPI_INFO_NULL, &fh ) ; MPI_File_set_view( fh, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; MPI_File_iwrite_at(fh, 10, &a, 1, MPI_INT, &reqs[0]) ; MPI_Wait(&reqs[0], &status) ; MPI_File_iread_at(fh, 10, &b, 1, MPI_INT, &reqs[1]) ; MPI_Wait(&reqs[1], &status) ;defines the same ordering as:
int a = 4, b; MPI_File_open( MPI_COMM_WORLD, "myfile", MPI_MODE_RDWR, MPI_INFO_NULL, &fh ) ; MPI_File_set_view( fh, 0, MPI_INT, MPI_INT, "native", MPI_INFO_NULL ) ; MPI_File_write_at(fh, 10, &a, 1, MPI_INT, &status ) ; MPI_File_read_at(fh, 10, &b, 1, MPI_INT, &status ) ;Since
Similar considerations apply to conflicting accesses of the form:
MPI_File_write_all_begin(fh,...) ; MPI_File_iread(fh,...) ; MPI_Wait(fh,...) ; MPI_File_write_all_end(fh,...) ;Recall that constraints governing consistency and semantics are not relevant to the following:
MPI_File_write_all_begin(fh,...) ; MPI_File_read_all_begin(fh,...) ; MPI_File_read_all_end(fh,...) ; MPI_File_write_all_end(fh,...) ;since split collective operations on the same file handle may not overlap (see Section Split Collective Data Access Routines ).