Our Master student Jhorman Jhair Gutiérrez Valderrama has just presented his Master Thesis titled GPU Performance and Consumption Analysis and Characterization for Graph Processing in a Heterogeneous Data Centers. The presentation took place in the School of Telecommunications Engineering, Technical University of Madrid (UPM). This Master Thesis is part of the Master program M.Sc. in Electronics Systems Engineering (MISE).
Big data applications have become increasingly demanding thanks to the boom of Cloud Computing. Massive graph processing algorithms used in social networks or web searchers are one of the Big data applications with higher demand. So, is of great importance to improve the efficiency of the Data Centers that process large scale applications such as graph processing algorithms. To achieve this, this Master Thesis presents a performance and consumption characterization of the Jetson TK1 development board from NVIDIA, with a 192 CUDA cores Tegra K1 GPU and 4 quad-core ARM Cortex-A15 processor. To characterize the board different type of graphs from real applications are used in conjunction with the PageRank algorithm to measure the level of importance of each node. Also, a comparison of performance and consumption is made with the execution of the graph algorithms on the quad-core ARM processor and on a high enterprise server.
The PageRank algorithm is implemented in the C++ programming language for the execution in the high enterprise server and the ARM processor, and in CUDA C++ for the execution in the Tegra K1 GPU. A monitoring system is implemented to gather temperature and consumption data from the Jetson TK1 development board. Also, with the help of software tools the total execution time and the PageRank processing time is gathered from the C++ and the CUDA C++ implementations. In order to select a representative set of graphs from the SNAP platform an estimation of the necessary memory for the algorithm execution is made.
Finally, by looking the behavior of the consumption, temperature and execution time for different type of graphs results showed different performance behavior as a function of the topology and type of graph. Lastly, different lines of research are proposed such as the characterization of other systems of computation that allows the execution of the PageRank algorithm.
Congratulations to Jhorman and his advisor Marina Zapater Sancho!