MPIPCL

Library implementation of MPI-4 Partitioned Communication (currently using persistent point-to-point communication).

This is a component of the MPI-Advance Project!

The API provided by this project can be found in mpipcl.h in /include/. The source code is in the /src/pt2pt folder. The current version of our library creates an individual MPI Request for each partition, that is stored in the MPIX_Request object. These partition requests are launched with the respective calls to MPI_PREADY (or its variants) on the send-side. On the receive-side, the partition requests are launched when the MPIX_Request is started. The current version of the library also features a progress thread in the background to synchronize the number of internal partitions to use (useful for when the send-side and receive-side partitions don't match).

The current implementation makes the following assumptions/shortcuts:

1. The datatype is the same on both the sender and receiver processes
2. Only contiguous datatype is supported
3. The amount of data described on the send-side must match the amount of data described on the receive-side, even if the partitions do not match
4. The status object is not updated; it is ignored
5. These API are invoked in MPI_THREAD_SERIALIZED mode (i.e., MPI_Init_thread with thread support for MPI_THREAD_SERIALIZED is used instead of MPI_Init)
6. A new MPI request object called MPIX_Request is used to support partitioned communication

Building

The MPIPCL library is a fairly simple cmake build:

mkdir <build_dir>
cd <build_dir>
cmake <options> ..
make

A loaded MPI package is required, along with CMake version 3.17 or higher. CMake should detect the MPI and pick the right compiler, but always double check the output to make sure.

Examples

There are six examples provided by this library (found in the examples/Basic folder) Each example has the specific execution instructions at the top, but most follow this pattern: mpirun -np <num_processes> ./<test_executable> <npartitions> <bufsize>

The first, second, and sixth example programs require only two processes, while the third, fourth, and fifth example programs require five processes. Note that bufsize % npartitions == 0 must be true for all examples and combinations of buffer sizes and number of partitions. The fifth test is testing all of the MPIX_Wait\Test functions and their variants.

If -DBUILD_EXAMPLES=ON is provided during the cmake command, the tests will also be built when building MPIPCL.