Coarray Fortran in CENTORI and a Sparse Matrix Multiplier Code

In modern high performance computers there is a trend towards a hybrid architecture. This hybrid architecture features distributed nodes, each of which contains multiple cores, over which memory is shared. A current challenge in HPC (High Performance Computing) is to find solutions which make the best possible use of hybrid architectures. One possible solution is the use of PGAS (Partitioned Global Address Space) languages. This project aims to investigate the potential of the PGAS language Co-array Fortran (CAF). The investigation is carried out by implementing CAF within key areas of two codes: a sparse matrix multiplier and the fusion code CENTORI. The merit of CAF was investigated by looking at two key aspects of the code produced: the code readability and maintainability, and code performance. A key part of the workings of an iterative solver for sparse linear systems is the multiplication of a sparse matrix by a vector. Code was created which parallelised this multiplication in CAF and OpenMP. The CAF parallelisation was done with both a shared vector and a distributed vector. The process of developing the CAF parallelisation is described. Comparison is made between the performance of the OpenMP parallelisation and the CAF parallelisations. This is done through benchmarks that try to mimic the usage of the sparse matrix multiplication in an iterative solver code. These benchmarks are carried out on the X2, XT4 and XT6 systems. In CENTORI the halo swap mechanisms were implemented in CAF. This was done both with and without buffers, taking advantage of the direct accessing capability of CAF. There were two versions created using buffers, one using single-sided put operations and one using single-sided get operations. Timers were added to the developed and original CENTORI codes, which gave the total time the program soent in halo exchange in each dimension. Benchmarks were carried out on the X2 and XT4 systems. In general the CAF code produced was found to be more readable than MPI implementations. This is further elaborated on within this report, along with the issues encountered in producing CAF implementations.