EMU: High-Performance Computing Center of SINAPAD

• Grant number: 23/07074-1
• Support Opportunities: Multi-user Equipment Program
• Start date: March 01, 2024
• End date: February 28, 2031
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Agreement: MCTI/MC
• Principal Investigator: Fabio Borges de Oliveira
• Grantee: Fabio Borges de Oliveira
• Host Institution: Laboratório Nacional de Computação Científica (LNCC). Petrópolis , SP, Brazil
• City of the host institution: Petrópolis
• Associated researchers: Aderson Farias do Nascimento ; Adriane Prisco Petry ; Alexandre Gonçalves Evsukoff ; Alvaro Luiz Gayoso de Azeredo Coutinho ; Amit Bhaya ; Anderson dos Reis Albuquerque ; André Rodrigues Muniz ; Antonio Luiz Pinho Ribeiro ; Antonio Tadeu Azevedo Gomes ; Argimiro Resende Secchi ; Carla Osthoff Ferreira de Barros ; Claysson Bruno Santos Vimieiro ; Djalma Mosqueira Falcão ; Fabio André Machado Porto ; Fabrício Rodrigues dos Santos ; Fernando Alves Rochinha ; Francisco Heron de Carvalho Junor ; Francisco Rodrigo Porto Cavalcanti ; Frederic Gerard Christian Valentin ; Hélio Anderson Duarte ; Janier Arias Garcia ; João Medeiros de Araújo ; Jose Dias do Nascimento Junior ; José Irineu Rangel Rigotti ; José Luis Drummond Alves ; Jose Miguel Ortega ; Kary Ann del Carmen Ocaña Gautherot ; Laurent Emmanuel Dardenne ; Leonardo Gregory Brunnet ; Luiz Manoel Rocha Gadelha Junior ; Marcio Arab Murad ; Marcio Rentes Borges ; Nelson Francisco Favilla Ebecken ; Nilo Sérgio Medeiros Cardozo ; Pablo Javier Blanco ; Paula de Miranda Ribeiro ; Philippe Olivier Alexandre Navaux ; Renato Antonio Celso Ferreira ; Roberto PInto Souto ; Rodrigo Affonso de Albuquerque Nóbrega ; Rodrigo Juliani Siqueira Dalmolin ; Rossana Maria de Castro Andrade ; Samuel Xavier de Souza ; Umberto Laino Fulco ; Von Braun Nascimento ; Willian Ricardo Rocha

Abstract

High Performance Processing (HPC) is currently a key area for scientific and industrial research. Broad-spectrum, high-cost, and computationally intensive research initiatives (such as high-energy physics, climatology, biomedicine, renewable energy, and aeronautics, to name a few) have historically promoted the area of HPC in more economically developed countries. This project aims to support the increase in the competitiveness of Brazilian scientific and industrial research through the acquisition/update of a set of 5 supercomputers in order to restore the computational capacity of the National High Performance Computing System - SINAPAD, instituted by Presidential Decree 5,156 of July 26, 2004. The project also includes the reformulation of SINAPAD's organizational model in tiers. In this new model, the LNCC, which coordinates SINAPAD by delegation from the MCTI through MCTI Ordinance No. 969 of 12/15/2006, will function as "Tier 0" with its supercomputer Santos Dumont (SDumont), with the aim of serving projects with demonstrated ability of their applications to scale from tens to hundreds of thousands of processors. The 5 SINAPAD centers participating in this proposal will work as "Tier 1", serving their regional communities for the execution of smaller projects with a demand for HPC services, projects that may eventually migrate to the LNCC when they demonstrate to scale to a supercomputer of larger size. For each center participating in this proposal, it is foreseen the adoption of a cluster architecture with the following general characteristics: number of computing nodes with only CPU varying between 10 and 40 (corresponding total number of CPU cores in each cluster varying between 1280 and 5120) ; number of compute nodes containing GPU cards ranging from 1 to 6 (corresponding total number of GPU cards in the cluster ranging from 8 to 24); RAM memory capacity per computational node ranging from 384 to 512 Gb (corresponding total memory ranging from 4 Tb to 18 Tb); storage capacity in a parallel file system shared between computational nodes ranging from 1.6 to 1.9 Pb; interconnect network of 200 Gbps (Infiniband). It is estimated, with these configuration profiles, an aggregate computational capacity of the 5 centers varying between 318 TFlops and 1.27 PFlops in CPUs and between 2.68 PFlops and 8.04 PFlops in GPUs, a capacity that will be added in SINAPAD to the current computational capacity of SDumont do LNCC, which is on the order of 4.1 PFlops (CPUs+GPUs). Here, the added capacity to be obtained in SINAPAD with GPUs is highlighted, being greater than that of SDumont and possibly also greater than IARA (supercomputer from SiDi dedicated exclusively to enterprise research in the area of AI). For the execution of this project, all participating centers make available teams experienced in the operation of supercomputers. (AU)