Tenable CUDA capabilities within the Tasmanian library, which allows for optimal performance on GPU accelerated architectures.
Significance and Impact
The bulk of the computational power of modern supercomputers derived from accelerators, either GPUs or Intel Phi. Toolkit for Adaptive Stochastic Modeling and Non-Intrusive ApproximatioN (TASMANIAN) is the ORNL flagship UQ library, enabling GPU acceleration is critical for allowing Tasmanian users to take advantage of modern accelerated architectures.
- Multi-platform build engine with large array of capabilities
- Consistent API independent from the available hardware architecture
Toolkit for Adaptive Stochastic Modeling and Non-Intrusive ApproximatioN (TASMANIAN) is an ORNL developed open source library for high dimensional integration and interpolation based upon sparse grids methods. Tasmanian offers the most feature rich sparse grids environment and is thus utilized by researchers in both industry and academic institutions. The two main UQ thrusts implemented in Tasmanian are the forward uncertainty propagation followed by the inverse problem for optimization and parameter calibration.
Sparse grids is a common method for creating computationally cheap surrogates to complex scientific
models when the models depend on multiple input variables. Studying the effects of uncertainty in each input parameter on the output of the model, i.e., uncertainty quantification, pertains to the statistical analysis of the input-output relationship and statistical methods rely heavily on Monte Carlo random sampling. While the Monte Carlo approach is very reliable, the convergence rate is slow and a huge number of samples is required to obtain statistically significant results. Such sampling would be prohibitive if applied directly to the full model, thus, cheap surrogates are used instead. Even after reduction, large number of samples collected from a sparse grid model with large number of outputs is computationally challenging. The current version of Tasmanian can leverage GPU accelerators to compute large batches of samples, achieving more than 5x speedup on OLCF Titan supercomputer.