Designing And Implementing A High-Performance Computing Heterogeneous Cluster

We present a new hybrid cluster, characterized by heterogeneous resources, set up in the Federico II University of Naples Data Center, funded by the IBiSCo (Infrastructure for BIg data and Scientific COmputing) project. It aims at big data analytics, high throughput and high performance processing,...

Full description

Saved in:

Bibliographic Details
Published in:	2022 International Conference on Electrical, Computer and Energy Technologies (ICECET) pp. 1 - 6
Main Authors:	Barone, Giovanni Battista, Bottalico, Davide, Carracciuolo, Luisa, Doria, Alessandra, Michelino, Davide, Pardi, Silvio, Russo, Guido, Sabella, Gianluca, Spisso, Bernardino
Format:	Conference Proceeding
Language:	English
Published:	IEEE 20-07-2022
Subjects:	Big Data CUDA Data centers Exascale computing File systems GPGPU HPC HPC for Big Data Analytics HTC Hybrid Computing Image processing Infini-Band Lustre Memory NVLink Parallel Computing Architecture RDMA Scientific computing UCX
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	We present a new hybrid cluster, characterized by heterogeneous resources, set up in the Federico II University of Naples Data Center, funded by the IBiSCo (Infrastructure for BIg data and Scientific COmputing) project. It aims at big data analytics, high throughput and high performance processing, image processing and analysis. The purpose of the hybrid features is to guarantee the best use of resources for their applications in different scenario, so as to profit from different computational paradigms: from parallel computing to GPGPU accelerated workload and their combinations. The cluster provides 128 GPUs as well as the coexistence of technologies for High Throughput Computing (HTC) and High Performance Computing (HPC). To offer heterogeneous resources, cluster nodes have multiple network connections together with an NVLink bus between the GPUs on each node, which ensures more efficient intranode communication. The data storage is separated from the computing nodes and its efficient access is assured by Lustre distributed and parallel file system which leverages on Infini-Band technology. Our work should be useful to evaluate some promising technologies for the management and the efficient usage of computing resources under development within different Exascale Computing Projects.
DOI:	10.1109/ICECET55527.2022.9872709