Optimization and Parallelization of a Time Series Classification Algorithm

Sinkovits, Robert S.

Permalink

https://hdl.handle.net/2142/104713 Copy

Description

Title

Optimization and Parallelization of a Time Series Classification Algorithm

Author(s)

Sinkovits, Robert S.

Contributor(s)

Huerta, Ramon

Issue Date

2019-08-20

Keyword(s)

• XSEDE
• Technical Report Series
• optimization
• parallelization
• MKL
• time-space tradeoff
• time series algorithm
• ECSS
• Extended Collaborative Support Services
• Ramon Huerta

Abstract

This technical report describes the steps taken to optimize and parallelize a time series classification algorithm as part of an Extended Collaborative Support Services (ECSS) project with XSEDE researcher Ramon Huerta at the University of California, San Diego. Switching from the GNU compiler to the Intel compiler and enabling Advanced Vector Extensions (AVX) resulted in a 2x speedup, while linking to the Intel Math Kernel Library (MKL) instead of the default LAPACK library further improved performance and provided an easy path to thread-level parallel execution. These changes resulted in a combined 46x speedup relative to a single core when running on all 16 cores of a dual-socket Intel Sandy Bridge node. Parallelization of several loops using OpenMP directives and the removal of an unnecessary duplicate call to a computationally demanding routine brought the total speedup to 86x. Optimization of linear algebra operations using time-space tradeoffs ultimately resulted in a total speedup of 168x relative to the original version and build of the code.

Publisher

Extreme Science and Engineering Discovery Environment (XSEDE)

Series/Report Name or Number

XSEDE-2019.1

Type of Resource

text

Language

eng

Permalink

http://hdl.handle.net/2142/104713

DOI

https://doi.org/10.21900/XSEDE-2019.1

Sponsor(s)/Grant Number(s)

National Science Foundation OCI-1053575

Owning Collections