High Performance Computing aspects of Single Particle Machine Learning

• Main Author: Näslund, Marcus
• Format: Others
• Language: English
• Published: Uppsala universitet, Institutionen för informationsteknologi 2015
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-260036

The vbSPT program is an existing MATLAB/C implementation of a variational Bayesian treatment of Hidden Markov Models to extract quantitative data from thousands of short single-molecule trajectories. In this work vbSPT is extensively profiled and optimized, including some attempts to parallelize using OpenMP. The underlying mathematical model is described in some detail and analyzed for future performance improvement. Results show that the previous parallelization scheme is inefficient and that optimization must be performed at a higher level than attempted here, which the report also details. The current implementation has very low potential for optimization, and this report recommends moving large parts of MATLAB code to C/C++, in part motivated by OpenMP offering a better speedup and scalability.