Programming Model Bake-off Helps Inform Data Analysis Directions in Post-Moore’s Era

Fig. 1: (left) reconstructed MicroCT data from the ALS undergoes oversegmentation (middle), then segmentation (right) using an advanced unsupervised learning method. Image courtesy T. Perciano, W. Bethel, C. Heinemann, D. Camp (LBNL).

Scientific Achievement

Detailed performance analysis of a state-of-the-art, unsupervised learning graphical model optimization method reveals new performance insights and contrasts between OpenMP, threaded, and data-parallel primitives (DPP) programming models [1].

Significance and Impact

All DOE mission science includes an element of data analysis that are challenged by large, complex data and processor architectures of increasing complexity. This work shows a way to obtain performance gains now in a platform portable way that holds promise for similar performance on future architectures.

Research Results

  • A parallel Markov Random Field graphical model optimization code is parallized using OpenMP, threads, and data parallel primitives.
  • Performance analysis measures multiple hardware performance counters and on multiple platforms.



  1. T. Perciano, C. Heinemann, D. Camp, B. Lessley, and E. W. Bethel, “Shared-Memory Parallel Probabilistic Graphical Modeling Optimization: Comparison of Threads, OpenMP, and Data-Parallel Primitives,” in High Performance Computing, Cham, pp. 127–145, Jun. 2020.