Brazilian university computing center participates in global program by Intel

The São Paulo State University is the first in the state to host a parallel computing center supported by the technology company.

By Elton Alisson

Agência FAPESP – Intel has selected the Center for Scientific Computing (CSC) at the São Paulo State University (Unesp) to become part of a parallel computing program, in which a large computational task is distributed and performed simultaneously by high-performance computers to optimize their use and reduce the time that it takes to complete the work.

Unesp will be the first university in São Paulo State and the second in Brazil to host one of the more than 40 Intel parallel computing centers throughout the world. Other centers are found at institutions such as Stanford University in the U.S., Eidgenössische Technische Hochschule Zürich (ETH) in Switzerland, Bristol University in the U.K., and the European Organization for Nuclear Research (CERN).

In Brazil, the company’s other parallel computing center is at COPPE – the Alberto Luiz Coimbra Institute – Graduate School and Research in Engineering at the Federal University of Rio de Janeiro (UFRJ).

According to Professor Sérgio Novaes, the scientific director of the Unesp CSC, the São Paulo university’s participation in the Intel program will allow both enhancement of the quality of services available for high-performance computation and support by GridUnesp, housed at the CSC, to nearly 60 scientific projects in various fields of research at the institution.

In particular, participation will allow training and development of CSC professionals in parallel programming techniques that will take maximum advantage of the new multi-core processors and accelerators installed at the center.

“It is critical that we not be left behind with regard to key advances in high-performance computing. We need to provide high-quality service to the Unesp academic community while, at the same time, monitoring technology developments in the field from up close, rather than superficially,” Novaes said.

The partnership will also allow application of the experience and knowledge acquired in parallel computing to experimental high-energy physics, which is one of the CSC’s fields of interest.

Careful selection

“The process of selecting the research institutions that become part of Intel’s parallel computing program is extremely rigorous, and few countries have two institutions selected to take part in the program, as Brazil does,” Fernando Martins, the president and general manager of Intel Brazil, told Agência FAPESP.

“This demonstrates not only the excellence of these institutions but also Brazil’s significance in the use and development of high-performance computing applications,” he emphasized.

According to Martins, one of the selection criteria for an institution wishing to participate in the program is having researchers with knowledge and experience in parallel computing interested in developing high-performance computing applications (software) in their fields of research excellence.

“Intel funds fellowships for those who will work in developing and enhancing the applications at these institutions,” Martins said.

The idea is that the applications developed at the parallel computing centers financed by the company be made available for use by the scientific and technical community in high-performance computing.

“The expectation is that the researchers and professionals who work on these software development projects share the results with the scientific and technical community in high-performance computing in Brazil through workshops organized for this purpose,” Martins said.

Computational dilemma

In addition to GridUnesp, the CSC at Unesp houses the FAPESP-funded São Paulo Research and Analysis Center (SPRACE). With this, the CSC provides specific computer support to the institution’s high-energy physics research group.

The computational structure of SPRACE is part of the Compact Muon Solenoid (CMS) collaboration on the Large Hadron Collider (LHC) operated by CERN.

“Through SPRACE, researchers and students from the Unesp high-energy physics group are using computer resources and storage installed at the CSC datacenter to process and to take part in analyzing the data generated by the CMS,” Novaes explained.

According to him, the scientific community involved in the experiments conducted using the LHC expect to face a “computational dilemma” starting in 2015.

This is because the LHC is scheduled to return to operation next year at higher-power levels, or between 13 and 14 tera electron volts (TeV), as opposed to the 8 TeV it has operated on in recent years. This power increase will be accompanied by an increase in the number of events at the collider (read more about this at http://agencia.fapesp.br/19825).

The problem, however, is that the installed computational capacity at the LHC, known as the Worldwide LHC Computing Grid (WLCG) – connects 100,000 processors in 40 countries, including those of SPRACE in Brazil – is not expected to be enough to process and store the enormous quantity of data that will be generated.

“Even from countries that are members of CERN, there is pressure to find a solution to this computational dilemma,” Novaes said.

“A natural solution would be to explore the new computer architectures that are coming onto the market, such as 256-core processors, to try to make the processes parallel and thus reduce the need for large investments in high-performance computing,” he explained.

The researchers are stepping up their efforts to develop and enhance the software to increase the parallelism and scalability of the applications.

In the case of CERN, one possible software candidate for conducting the parallel computing required for the tremendous amount of LHC data starting in 2015 is the Geant4, developed by a group of high-energy physics researchers at CERN.

The software allows simulation of radiation interactions with matter, as occurs inside the particle detectors of the LHC. “The Geant4 simulates practically all of the events that involve the interaction of radiation with matter that can occur inside the detector.”

“It is through this simulation that we will be able to predict how the particles detected in the collider will be measured,” Novaes explained. Because the interaction of radiation with matter is of interest beyond high-energy physics, the software application is beginning to expand to other fields, such as radiation therapy.

“Today, the software is used to calculate the dosimetry of radiation for treating cancer when the dose of radiation has to be focused on a target tumor in order to reduce the exposure of healthy tissue,” Novaes said.

According to the researcher, under the scope of the Intel parallel computing program, the CSC’s goals will be to adapt the software to modern computing architectures and to develop new strategies and algorithms so that the program makes more effective use of the new multi-core processors.

Additionally, the researchers expect to continue their collaborative efforts with colleagues from the simulation group at the Fermi National Accelerator Laboratory (Fermilab) in the United States, dedicated to optimizing the Geant4 and working in partnership with software engineers from CERN to develop a new version of the software: GeantV.

“The project will enable the selection of at least two people to work full-time on parallelization of the Geant,” Novaes said.

“The idea is that the selected professionals will remain abroad for at least three months, working with researchers at Fermilab, CERN and Intel, to become involved in developing parallel software before returning to Brazil to continue the work,” he said.