Ichitaro Yamazaki, Jennifer Loe, Christian Glusa, Sivasankaran Rajamanickam Piotr Luszczek, and Jack Dongarra
HPG-MxP is a software package that performs a fixed number of multigrid preconditioned (using a Gauss-Seidel smoother) Generalized minimal residual (PGMRES) iterations.
The HPG-MxP rating is a weighted GFLOP/s (billion floating-point operations per second) value that is composed of the operations performed in the PGMRES iteration phase over the time taken to finish the solve. The overheads such as the time of problem construction and any structural modifications to improve the final performance are amortized over several iterations (the amortization weight) and added to the runtime.
Integer arrays describing the original problem structure have global and local scope (the global indices are unique across the entire distributed memory system, while the local indices are unique only within a single memory image of a single MPI rank). Integer data for global/local indices have three modes:
These various modes are required in order to address sufficiently big problems if the range of indexing exceeds value of 2 to power 31 (roughly 2.1B), or to conserve storage costs if the range of indexing is less than 2 to power 31.
The HPG-MxP software package requires the availibility on your system of an implementation of the Message Passing Interface (MPI) if MPI build of HPG-MxP was enabled at configuration time, and a compiler that supports OpenMP pragma syntax. An implementation compliant with MPI version 1.1 is sufficient.
HPG-MxP can be run in just a few minutes from start to finish. However, official runs must be at least 1800 seconds (30 minutes) as reported in the output file. The Quick Path option is an exception for machines that are in production mode. In this situation, which should be confirmed by sending a note to the HPG-MxP Benchmark authors, the Quick Path option can be invoked by setting the run time parameter equal to 0 (zero).
A valid run must also execute a problem size that is large enough so that data arrays accessed in the GMRES iteration loop do not fit in the cache of the device in a way that would be unrealistic in a real application setting. At present, this restriction means that the problem size should be large enough to occupy a significant fraction of the main memory, or at least 1/4 of the total.
Future memory system architectures may require restatement of the specific memory size requirements. But the guiding principle will always be that the problem size should reflect what would be reasonable for a real sparse iterative solver.
The source code documentation can be generated with a Doxygen (version 1.8 or newer).