Creating and sharing knowledge for telecommunications

Clustering Low-Cost, Cloud-Based Servers to Solve Intensive, Parallel Computations

Brás, N. ; Valadão, G.

Clustering Low-Cost, Cloud-Based Servers to Solve Intensive, Parallel Computations, Proc Conf. on Telecommunications - ConfTele, Aveiro, Portugal, Vol. -, pp. - - -, September, 2015.

Digital Object Identifier:

Download Full text PDF ( 235 KBs)

This paper advocates the usage of available cloud based services for intensive parallel floating point computations, based on clusters of servers installed with Graphical Processing Units (GPUs), in order to run low-cost, High Performance Computing (HPC) tasks.
It is described a cluster of multiple servers installed with GPU units and running open-source software which works as an easy to scale, low cost platform with centralized Master-Slave control, able to turn on/off the Slaves (GPU server machines) as needed.
The objective is to show that, nowadays, parallelization ar- chitectures allow to decrease dramatically the computing time while maintaining or even reducing the intrinsic costs (e/hour) comparing with traditional approaches taken by researchers and industry practitioners.
A computer vision classification problem is used as a toy problem: a deep neural network training using a standard image data library, through known and easy to use parallelization packages. For this problem, comparisons on time performance, scalability and costs, with regard to other traditional approaches, are presented. In this particular case, the computing time was reduced almost 30 times while attaining however, an unexpected cost reduction.