The PARKBENCH suite includes five matrix kernels:
  1. Dense matrix multiply. Communication involves broadcast of data along rows of mesh, and periodic shift along column direction (or vice versa).
  2. Transpose. Matrix transpose is an important benchmark because it exercises the communications of computer heavily on a realistic problem where pairs of processors communicate with each other simultaneously. It is a useful test of the total communications capacity of the network.
  3. Dense LU factorization with partial pivoting. Searching for a pivot is basically a reduction operation within one column of the processor mesh. Exchange of pivot rows is a point-to-point communication. Update phase requires data to be broadcast along rows and columns of the processor mesh.
  4. QR Decomposition. In this benchmark parallelization is achieved by distribution of rows on a logical grid of processors using block interleaving.
  5. Matrix tridiagonalization, for eigenvalue computations of symmetric matrices.
These kernels may be obtained in the current distribution from the netlib repository.

PARKBENCH kernels page

Last Modified May 14, 1996