Heterogeneous Computing for Markov Models in Big Data

Abstract

Many Big Data problems, Markov Model related included, are solved using heterogenous systems: host + parallel programmable accelerator. The current solutions for the accelerator part - for example, GPU used as GPGPU - provide limited accelerations due to some architectural constraints. The paper introduces the use of a programmable parallel accelerator able to perform efficient vector and matrix operations avoiding the limitations of the current systems designed using "of-theshelf" solutions. Our main result is an architecture whose actual performance is a much higher percentage from its peak performance than those of the consecrated accelerators. The performance improvements we offer come from the following two features: the addition of a reduction network at the output of a linear array of cells and an appropriate use of a serial register distributed along the same linear array of cells. Thus, for a n-state Markov Model, instead of a solution with the size in O(n2) and an acceleration in O(n2=logn), we offer an accelerator with the size in O(n) and the acceleration in O(n).