The ext_mpi_collectives library provides persistent collective communication operations according to the MPI 4.0 standard. It is built on top of existing MPI libraries. If the underlying library is MPI 4.0 enabled the MPI profiler hook is applied for linking. The fast execution of the collectives through the library is provided by the following features:
- Communication with flexible number of ports per node depending on the message size
- Benchmark at installation time of the library for latency and bandwidth measurements
- Cyclic shift algorithm for flexible number of ports for allgatherv, reduce_scatter and allreduce
- Rank reordering for variable message sizes of allgatherv and reduce_scatter
- Extension of cyclic shift algorithm with prefix sum for allreduce
Clone the repository:
git clone https://github.com/eth-cscs/ext_mpi_collectives.git
cd ext_mpi_collectives
If necessary, adapt the MPI wrapper of the C compiler in the first line of the Makefile. Then just run:
make
If the existing MPI library supports the standard 4.0 it is just required to re-link the application with the library
mpicxx main.o -Lext_mpi_path/lib -lext_mpi_collectives -o executable.x
if not the header file ext_mpi_path/include/mpi/ext_mpi_interface.h needs to be included in the files calling the libraries and the application needs to be recompiled, e.g.
mpicxx main.cc -Lext_mpi_path/lib -lext_mpi_collectives -o executable.x
When compiling the library also a benchmark is generated which serves also as example for the library's usage
mpirun -n 36 ./bin/benchmark.x
The library supports CUDA aware MPI using this feature from the communication primitives of the underlying library. For using it the alternative Makefile
make -f Makefile_gpu
is provided with the benchmark
mpirun -n 36 ./bin/benchmark_gpu.x
This feature is not highly optimised.
Only the collective operations MPI_Allreduce_init, MPI_Allgatherv_init, MPI_Reduce_scatter_init, MPI_Bcast_init, MPI_Reduce_init, MPI_Gatherv_init, MPI_Scatterv_init, MPI_Allgather_init, MPI_Reduce_scatter_block_init, MPI_Gather_init, MPI_Scatter_init are provided. Datatypes need to be contiguous and communicators need to be intra communicators. For reduction operations only the datatypes float, double, int and long int with the operation MPI_SUM are supported.
The components of the library which do not require MPI can be tested independently. Every of the following executables takes the input of the previous one and further processes the data. The different steps represent a mini language and an assembler code which is optimised step by step.
./bin/get_input_allreduce.x | ./bin/test_allreduce.x | ./bin/test_raw_code_tasks_node.x | ./bin/test_reduce_copyin.x | ./bin/test_raw_code.x | ./bin/test_reduce_copyout.x | ./bin/test_buffer_offset.x | ./bin/test_no_offset.x | ./bin/test_optimise_buffers.x | ./bin/test_optimise_buffers2.x | ./bin/test_parallel_memcpy.x | ./bin/test_raw_code_merge.x
There is also a standalone test for automatic parameter detection of the allreduce algorithm
./bin/simulate.x
- Required before running cmake:
./minclude.sh latency_bandwidth/ext_mpi_bm.txt > latency_bandwidth.tmp
cat latency_bandwidth/ext_mpi_nst.txt > node_size_threshold.tmp
- Build libext_mpi_collectives.a with CMake:
cmake -B build -S ext_mpi_collectives.git \
-DBUILD_TESTING=ON
cmake --build build -j
# cmake --build build -t help
# cmake --build build -t test_waitany.x -v
cmake --install build --prefix myinstalldir
# -- Installing: ./myinstalldir/lib/libext_mpi_collectives.a
- Build with:
cmake -B build+cuda -S ext_mpi_collectives.git \
-DBUILD_TESTING=ON -DCMAKE_CUDA_ARCHITECTURES=80
cmake --build build+cuda -j
cmake --install build+cuda --prefix myinstalldir
A. Jocksch, N. Ohana, E. Lanti, V. Karakasis, L. Villard: Optimised allgatherv, reduce_scatter and allreduce communication in message-passing systems. poster at PASC 2021
A. Jocksch, N. Ohana, E. Lanti, V. Karakasis, L. Villard: Towards an optimal allreduce communication in message-passing systems. poster at EuroMPI/USA'20
A. Jocksch, M. Kraushaar, D. Daverio: Optimised all-to-all communication on multicore architectures applied to FFTs with pencil decomposition. Concurrency and Computation: Practice and Experience, Vol 31, No 16, e4964, 2019