Fast Multidimensional Performance Parameter Estimation with Multiple One-Dimensional d-Spline Parameter Search

One approach of software automatic tuning is the incremental performance parameter estimation (IPPE) method, which is based on discrete spline function (d-Spline). We have applied IPPE to multidimensional performance parameter estimation. IPPE starts with the minimum number of the sampling points. It adds other sampling points by updating d-Spline at every iteration. For multidimensional performance parameter estimation, IPPE method using multidimensional d-Spline was studied in our previous work. However, the parameter estimation cost increases significantly as the number of performance parameters increases, because the calculation cost of multidimensional d-Spline is much higher than one-dimensional d-Spline. Computational cost for one-dimensional d-Spline is negligible. In this study, we propose a method for estimating the optimum value of multidimensional performance parameters by repeating one-dimensional d-Spline parameter search. We demonstrated that our method is able to identify optimal parameters with little cost. In case of three dimensional performance parameter estimation, our proposed method required only 104.8 combinations out of 1440 possible parameter combinations. We experimentally showed that multiple one-dimensional d-Spline parameter search is efficient for multidimensional performance parameter estimation.

[1]  R. Franke.  A Critical Comparison of Some Methods for Interpolation of Scattered Data , 1979 .

[2]  Yuefan Deng,et al.  New trends in high performance computing , 2001, Parallel Computing.

[3]  Reiji Suda,et al.  A Bayesian Method for Online Code Selection : Toward E ffi cient and Robust Methods of Automatic Tuning , 2007 .

[4]  Takahiro Katagiri,et al.  Enhancement of Incremental Performance Parameter Estimation on ppOpen-AT , 2015, 2015 IEEE 9th International Symposium on Embedded Multicore/Many-core Systems-on-Chip.

[5]  Chi-Bang Kuan,et al.  Automated Empirical Optimization , 2011, Encyclopedia of Parallel Computing.

[6]  Takahiro Katagiri,et al.  Auto-tuning of Computation Kernels from an FDM Code with ppOpen-AT , 2014, 2014 IEEE 8th International Symposium on Embedded Multicore/Manycore SoCs.

[7]  Takahiro Katagiri,et al.  d-Spline Based Incremental Parameter Estimation in Automatic Performance Tuning , 2006, PARA.

[8]  Marian Brezina,et al.  Algebraic multigrid by smoothed aggregation for second and fourth order elliptic problems , 2005, Computing.

[9]  James Demmel,et al.  Optimizing matrix multiply using PHiPAC: a portable, high-performance, ANSI C coding methodology , 1997, ICS '97.

[10]  Satoshi Ito,et al.  Early Experiences for Adaptation of Auto-tuning by ppOpen-AT to an Explicit Method , 2013, 2013 IEEE 7th International Symposium on Embedded Multicore Socs.

[11]  Takahiro Katagiri,et al.  Implementation of d-Spline-based incremental performance parameter estimation method with ppOpen-AT , 2014, Sci. Program..