Introduction to High-Performance Computing Using MPI - High-Throughput Image Reconstruction and Analysis - page 134

Asynchronous Communication
The functions MPI Send and MPI Recv provide the functionality of a blocking
communication. The MPI Send call blocks until the data has either been sent on
the network or copied into the system buffers. The application is free to modify
the buffer given as the input argument to the function MPI Send , as soon as the
call returns. Similarly, the MPI Recv function call blocks until the required data
is available and copied into the receive buffer specified in the call.
For performance reasons and to obtain better overlap of computation
and communication, it is desirable to have functions using which the ap-
plications can initiate a communication asynchronously, monitor its progress,
check the availability of the data, and get information about the state
of the network. Several asynchronous communication functions are avail-
able to achieve this goal. These functions include MPI Isend , MPI Ibsend ,
MPI Issend ,and MPI Irsend for different flavors of asynchronous send opera-
tions, MPI Irecv for asynchronous receive operation, MPI Test , MPI Testany ,
MPI Testall ,and MPI Testsome for monitoring the progress of a previously
initiated communication operation, MPI Wait , MPI Waitany , MPI Waitall ,
and MPI Waitsome for blocking until the specified communication operation is
complete, MPI Probe for checking if there is a message ready to be delivered to
the process, and MPI Cancel for canceling a previously initiated asynchronous
communication.
Collective Communication Operations
In addition to MPI Bcast and MPI Allreduce , several other collective commu-
nication operations are available in MPI. The function MPI Gather is used to ag-
gregate data from all the processes of a communicator into a single large buffer of a
specified root process. The scatter operation, using the function MPI Scatter ,is
the opposite. It slices an array stored at a root process and distributes it to the other
processes in the communicator. The functions MPI Gatherv and MPI Scatterv
may be used when the size of the data at different processes is different. The
functions MPI Allgather and MPI Allgatherv are variants of MPI Gather
and MPI Gatherv , where the data collected at the root process is simultaneously
broadcast to all the processes. The functions MPI Alltoall and MPI Alltoallv
are used for more general communication pattern, in which each node has data
for every other node. For example, the function MPI Alltoall may be directly used
for the transpose step of parallel FFT implementations [25].
MPI-IO
The MPI-2 specifications [28] provide a natural approach to file I/O that is consis-
tent with the overall design of MPI. It supports basic file operations such as open,
close, read, write, delete, and resize, in addition to enhanced I/O capabilities.
The concept of the view of a file is central to the design of MPI-IO. It is defined
as a repetitive pattern of noncontiguous blocks of the file data that is accessible and
appears contiguous to a process. A view is created using a displacement ,an etype ,
and a filetype . An etype is an elementary unit of transfer in MPI-IO. All the file I/O
is done in units that are multiples of an etype. An etype may be defined as a fixed