ROMJIST Volume 25, No. 3-4, 2022, pp. 368-380
George-Vlăduț POPESCU Improvements in Data Transfer for a MapReduce Accelerator
ABSTRACT: This paper presents a way to improve data transfer for a MapReduce Accelerator designed for parallel data processing and integrated into a heterogeneous computing system. The proposed solution is a new module called the Data Transfer Engine, integrated into the structure of the Accelerator. This module takes over the tasks related to data transfer, separating the data processing and data transfer flows, thus reducing the total execution time of a program. Two algorithms for processing large matrices were proposed to evaluate the improvement in the total execution time brought by the Data Transfer Engine. These algorithms are based on a custom linear algebra library used to run basic matrix operations on the target Accelerator.KEYWORDS: Data Transfer Engine; heterogeneous computing; improved data transfer; MapReduce AcceleratorRead full text (pdf)
