As parallel computing tends toward the exascale, scientific data produced by simulations are growing increasingly massive, sometimes resulting in terabytes of data.  By viewing this data as a dense tensor, we can compute a Tucker decomposition to find inherent low-dimensional multilinear structure, achieving impressive compression ratios with negligible loss in accuracy.  We present recent improvements in our distributed-memory parallel implementation of the Tucker decomposition, whose key computations correspond to parallel linear algebra operations.  To demonstrate the compression and accuracy of the method, we apply our software to real-world data sets from combustion simulations.  We also provide detailed performance results.

Bio:
Dr. Alicia Klinvex received her bachelors degree in computer science at the Pennsylvania State University. She earned a PhD in computer science at Purdue University, where she studied under Dr. Ahmed Sameh. Her thesis topic was scalable symmetric eigensolvers. Alicia is a postdoctoral researcher at Sandia National Laboratories, where she primarily works on the IDEAS software productivity project. Her other projects include spectral clustering for graphs and hypergraphs, and tensor decompositions.