Parallel Halo Exchange Using Fortran Coarrays

This repository contains code and test cases for investigating Fortran coarray implementations of a parallel halo exchange operation associated with domain decomposition methods for PDE. Here the domain is partitioned into P subdomains, one for each of P processes (or images), and each process calculates the unknown variables defined on its subdomain. Processes need to exchange values along their subdomain boundary with their neighbors in a halo exchange operation.

Background

The particular form of the halo exchange considered here originates from the MPI-parallel code Truchas. Abstractly one starts with a (global) index set ${1, 2, ..., N}$. Given $P$ processes, an index is assigned to a unique process according to a block partition of the index set: the first $n_1$ indices are assigned to process 1, the next $n_2$ to process 2, etc. An index so assigned to a process is said to be owned by that process. In addition to these, a process may include indices owned by other processes. The owned indices are said to be on-process and these additional indices, if any, are said to be off-process. For the purposes of computation, the collection of all indices known to a process are mapped to a process-local index set: the block of on-process indices are mapped, in order, to consecutive local indices starting at 1, and the off-process indices are mapped to consecutive local indices immediately following these.

Now the elements of an array indexed by this index set will be distributed across the $P$ processes according to the distribution of the indices. Each global index is owned by exactly one process (where it is on-process) but may exist on other processes as an off-process index. The mapping from off-process indices to their corresponding on-process index is a many-to-one mapping. A fundamental operation on the distributed array is to replace the value at each off-process index with the value at its corresponding on-process index; this is referred to as a gather (of off-process data) operation.

Coarray Implementations

In the gather (or halo exchange) operation each image sends some elements of its on-process data to other images, where it is received to overwrite off-process data, and likewise receives on-process data from other images to overwrite its own off-process data. There are 4 different coarray implementations of this operation. They all perform exactly the same data exchange, differing only in how it is organized.

• Method 1: Scattered read from remote image Each image reads indirectly-indexed on-process data from remote images to fill its (contiguous) array of off-process data
• Method 2: Blocked-read from remote image Each image pre-gathers its on-process data destined for other images into contiguous blocks of a send buffer, and then each image reads its buffer blocks from remote images to fill corresponding blocks of its off-process data array.
• Method 3: Gathered write to remote image Each image writes its indirectly indexed on-process data to (contiguous) blocks of remote images off-process data.
• Method 4: Blocked-write to remote image Each image pre-gathers its on-process data destined for other images into contiguous blocks of a send buffer, and then writes those blocks to corresponding blocks of remote image off-process data arrays.

Method 1 is the most straightforward and simple implementation. The other methods were intended to explore whether there is a performance difference between reading from or writing to remote images, and whether performance gains could be made by structuring the transfers to/from remote images in contiguous blocks of data.

The distributed index set mapping and the associated gather operation are implemented by the module index_map_type whose source is found in the coarray/method* directories for the different methods. The gather operation is performed by the module procedure gather_aux, and the configuration of the communication pattern used by that procedure is generated by the module procedure add_offp_index.

Note that a fuller-featured, production index_map_type module with both MPI and coarray implementations can be found at https://github.com/nncarlson/index-map.

Your comments are very much welcome; creating an issue in the issue tracker is the best way to provide feedback.

Reference MPI Implementation

An MPI implementation of the halo exchange is found in the mpi directory. This serves as a baseline against which to assess the different coarray implementations. It uses a graph communicator and MPI-3 neighborhood collective to perform the halo exchange.

The Tests

The file main.f90 is a test driver. It reads data that describes the partitioning of the index set and then performs the gather operation on an integer array. The on-process elements of the array are initialized with their corresponding global IDs and the off-process elements with invalid data (-1). After the gather operation the off-process elements should be filled with their global IDs, and this is checked. To get more accurate timings the gather operation may be repeated multiple times, which is especially important for the smaller datasets.

The test data is stored in subdirectories of the test-data directory, one for each test. A subdirectory contains a collection of input files, one per image. Each file (unformatted stream) consists of 2 records. The first consists of the block size assigned to the image (i.e., the number of on-process indices) and the number of off-process indices. The second record is the global IDs of the off-process indices in strictly increasing order.

The current test data was generated by a version of Truchas hacked to output this internal data. It comes from an unstructured finite element type mesh partitioned using METIS and corresponds to the cell index set (there are also the node, face, and edge index sets that could be obtained). There is data from a series of meshes ("opencalc-B") of increasing size:

Mesh	B0	B1	B2	B3	B4	B5
Cells	70K	206K	562K	1.6M	4.4M	13.4M

Each mesh is partitioned into various numbers of partitions. The mesh and number of partitions is reflected in the name of the test subdirectory.

Compiling

The project uses CMake (version 3.22 or later) to compile the tests. You may need to set your FC environment variable to the name or path of your Fortran compiler before running cmake to ensure that CMake finds the correct compiler. The CMake setup understands how to compile coarray code when using one of the following Fortran compilers:

• NAG 7.1 or later with its built-in coarray support on shared-memory systems.

• Intel OneAPI with its built-in coarray support. Both the classic ifort and and new LLVM-based ifx compilers are supported. The companion Intel MPI package must be installed and Intel's setup script run to configure your environment. The Intel coarray implementation uses MPI under the hood.

• GFortran with OpenCoarrays. OpenCoarrays supplies the implementation of coarrays used by the gfortran compiler. Be sure the bin directory of the opencoarrays installation is in your path so that the compiler wrapper caf and runner cafrun can be found. Set FC=caf before running cmake. OpenCoarrays uses MPI under the hood, and at the time of this writing is compatible with MPICH version 4.0, but not 4.1 or later. Refer to the OpenCoarray website linked above for requirements.

To clone the repository and compile the tests:

$ git clone https://github.com/nncarlson/coarray-halo-exchange.git
$ cd coarray-halo-exchange
$ mkdir build
$ cd build
$ cmake .. # cmake options go here
$ make

Optimized tests will be built by default using CMake's default flags for the "Release" build type and your specific compiler. Compiler flags can be set explicitly on the cmake command line by defining the CMAKE_Fortran_FLAGS variable; e.g. -D CMAKE_Fortran_FLAGS="-O3"

To build the MPI version of the test use this cmake command line instead:

$ cmake .. -D BUILD_MPI_TEST=YES

The test executables will be found in build/coarray (or build/mpi).

Running

The test executables take 1 or 2 command line arguments. The first is the path to the directory containing the data files for the test. The second is the number of times to repeat the gather operation before collecting timing data. If not specified it defaults to just 1. Only the gather operation itself is timed, and the average time per gather call is reported.

Here's an example of how to run the coarray1 test from the build/coarray directory using the small "B0" dataset and 4 coarray images, averaging the time for the gather operation over 1000 iterations:

• Intel OneAPI:

  $ FOR_COARRAY_NUM_IMAGES=4 ./test-coarray1 ../../test-data/opencalc-B0-4 1000
  

• GFortran/OpenCoarrays:

  $ cafrun -n 4 ./test-coarray1 ../../test-data/opencalc-B0-4 1000
  

• NAG:

  $ NAGFORTRAN_NUM_IMAGES=4 ./test-coarray1 ../../test-data/opencalc-B0-4 1000
  

Results

• Linux / 32-core Threadripper 7970X / 32 images
• Linux / 12-core Threadripper 2920X / 12 images
• Variations of version1; NAG beats MPI