Time Series Heterogeneous Co-execution on CPU+GPU

Abstract

Time series motif (similarities) and discords discovery is one of the most important and challenging problems nowadays for time series analytics. We use an algorithm called “scrimp” that excels in collecting the relevant information of time series by reducing the computational complexity of the searching. Starting from the sequential algorithm we develop parallel alternatives based on a variety of scheduling policies that target different computing devices in a system that integrates a CPU multicore and an embedded GPU. These policies are named Dynamic -using Intel TBB- and Static -using C++11 threads- when targeting the CPU, and they are compared to a heterogeneous adaptive approach named LogFit -using Intel TBB and OpenCL- when targeting the co-execution on the CPU and GPU.