Central Research Laboratory, Hitachi Ltd.
"Numerical Policy Interface for Automatic Tuning Library"
Recently, parameters for performance tuning are diverse in many matrix
libraries, thus requiring automatic tuning facilities. Most of the existing
researches focus on computation time reductions, while many matrix library
users' requirements are becoming hybrid, due to today's high performance
platforms complexity. For example, a user wants to reduce the computation
time under the limitation of 10 gigabyte memory use. In this talk, recent
research trends in automatic tuning technology are reviewed and a new framework
"numerical policy" is proposed to develop high-quality matrix
libraries. The framework is applied to Lanczos eigensolver on an SMP platform,
one node of an SR11000. The result shows that the framework effectively
selects parameter values that achieve a good balance between memory requirements
and computation time. The result implies that "numerical policy"
framework plays an important role to solve the diversity problem in selecting
the tuning parameter values in matrix libraries.