International Workshop on Open MP (IWOMP)

Meng Zhang, Alice Koniges, Yingjie He, Samuel Williams, Thorsten Kurth, Brandon Cook, Jack Deslippe, Al BertozzieScholarship 2017

Abstract

We investigate the OpenMP parallelization and optimization\nof two novel data classification algorithms. The new algorithms are\nbased on graph and PDE solution techniques and provide significant accuracy and performance advantages over traditional data classification\nalgorithms in serial mode. The methods leverage the Nystrom extension\nto calculate eigenvalue/eigenvectors of the graph Laplacian and this\nis a self-contained module that can be used in conjunction with other\ngraph-Laplacian based methods such as spectral clustering. We use performance\ntools to collect the hotspots and memory access of the serial\ncodes and use OpenMP as the parallelization language to parallelize the\nmost time-consuming parts. Where possible, we also use library routines.\nWe then optimize the OpenMP implementations and detail the performance\non traditional supercomputer nodes (in our case a Cray XC30),\nand test the optimization steps on emerging testbed systems based on Intel’s\nKnights Corner and Landing processors. We show both performance\nimprovement and strong scaling behavior. A large number of optimization\ntechniques and analyses are necessary before the algorithm reaches\nalmost ideal scaling.
BibTeX
@inproceedings{zhang2017international,
  author = {Meng Zhang and Alice Koniges and Yingjie He and Samuel Williams and Thorsten Kurth and Brandon Cook and Jack Deslippe and AL Bertozzi},
  title = {{International Workshop on Open MP (IWOMP)}},
  booktitle = {eScholarship},
  year = {2017},
}

Loading related content...

Figure