|
Title:
|
MINIMIZING DATA TRANSFERS ON MATRIX MULTIPLICATIONS IN GPU-BASED HETEROGENEOUS ENVIRONMENTS |
|
Author(s):
|
Ricardo I. A. e Silva, Tiago Sotana, Rodolfo M. de Barros, Jacques D. Brancher |
|
ISBN:
|
978-989-8533-14-2 |
|
Editors:
|
Hans Weghorn and Pedro Isaías |
|
Year:
|
2012 |
|
Edition:
|
Single |
|
Keywords:
|
Matrix Multiplication, Heterogeneous, GPGPU, Parallelism, Memory Transfer |
|
Type:
|
Full Paper |
|
First Page:
|
187 |
|
Last Page:
|
193 |
|
Language:
|
English |
|
Cover:
|
|
|
Full Contents:
|
click to dowload 
|
|
Paper Abstract:
|
Heterogeneous environments in computation systems, composed by a CPU and one or more GPUs, present opportunity in trying to increase performance of linear algebra algorithms. If computational capabilities of each available processor are summed, it can reach respectable few teraFLOPS (TFLOPS). But for it to be possible, a task and its dataset must be parallelized and distributed. Also, data transfers between processors memory must be carefully considered, since they are slow and can hinder global performance. Matrix multiplication is a common task of such systems. It is a highly parallelizable operation and useful, since it is present in most linear algebra algorithms. But its utilization can fall short of performance with processors heterogeneity because of the required high amount of memory transfers. In this work we present a method of distributing tasks and dataset from matrix multiplications that minimizes memory transfers, while trying to utilize 2 GPUs and the CPU. We show that we minimize time spent with transfers from CPU to GPU down to 60% in comparison to execution in single GPU, for very large matrices cases. |
|
|
|
|
|
|
