2021-03-03-Person to watch: Bengisu Elis

They are researching future technologies such as quantum computers, developing innovative technology or algorithms, advancing supercomputing, artificial intelligence or machine learning: In loose succession, we will introduce young scientists here, from whom we are sure to hear more. Second person is bengisu Elis, who researches for the future of supercomputers at Technical Univeristy Munich.

beast

Research for supercomputing

It's not easy to describe her job. "When computer units are connected with a cable, they communicate with each other when computing. My job is to improve this communication," says Bengisu Elis. This is how she usually explains to her parents – both bank employees and specialists in financial investments –what is behind her research work for her PhD thesis in computational science. The PhD student at the Technical University of Munich (TUM) explains her work to her younger brother using practical examples. She talks about research projects for which thousands of processors are assembled into supercomputers to calculate measurement data and use it to simulate weather conditions, the air currents around cars and railways or to evaluate the properties of materials and chemical compounds.

Technical assistance for research and science

"Supercomputers are not meant for the general public, but they sometimes work for days or weeks for a research result," says Elis. "My job in this is to improve performance and runtimes to save time and sometimes money and give researchers better ways to answer questions." In addition to the Linux cluster and the SuperMUC-NG of the Leibniz Supercomputing Centre (LRZ), the BEAST test environment is now also one of Bengisu Elis' work environments: "The new test bed is technically interesting and an extremely important tool for my work; I can use it to investigate different architectures." Therefore Elis supports the BEAST practical course as a lecturer and supervises experiments or students' Bachelor's theses, but especially for her PhD she is testing how to optimise the communication of processor types and thus speed up their work. At this point, the talk quickly turns to the "beauty of computational sciences", to the meaning and motivation of dealing with control units, processors, chips and supercomputers in the first place: "In many areas of the natural and life sciences, experiments are not possible, too expensive or cumbersome. Computational sciences and high-performance computing help here with technology and algorithms to pursue research questions. That's wonderful," says Elis with a short laugh.

Enthusiasm resounds, but instead of raving about new technologies and opportunities, Elis prefers to engage with arguments and explanations. In conversation, she listens carefully, thinks longer about some answers, underlines them with certain, calm gestures. She is straightforward, very likeable, hands-on, optimistic: "Due to the growing amount of data, computational sciences are penetrating every area of knowledge, which is why research in this field is so important. That makes me happy, I solve problems and develop ideas for future IT architectures that can help science and society."

Optimising the communication of processors for more performance

The young researcher came to her vocation rather by chance and at the end of her electrical engineering studies at Orta Doğu Teknik Üniversitesi in Ankara. In 2015, she examined antenna and chip technology in a volunteer research project, modelled functions on the computer and was immediately fascinated. At the defense company Aselsan, she deepened her experience in Electronic Engineering and High-Performance Computing (HPC), but left the company after a short time, went to Germany and took up a Master's degree in Computational Sciences at the Technical University of Munich (TUM): "HPC is supposed to help humanity, I was more interested in the human side." Since 2019, Elis has been working on her PhD, for which she studies the Message Passing Interface (MPI), a standard for communication between processors or control units. This is probably where the near future of supercomputing will be decided.

The number of processors that can be switched in parallel is gradually reaching its limits. The systems consume a lot of power, and their tasks are also changing due to growing data volumes. Take SuperMUC-NG, for example: here, more than 311,000 processors can work simultaneously, requiring up to four megawatt - about as much as a four-person household in a year. Improved communication between the control units (CPU) could speed up work and help reduce power requirements. The effect is quite similar when CPUs are combined with specialised control units, for example graphics processing units (GPUs), and/or so-called accelerators. However, this not only requires a new supercomputer design, but also changes communication. Consequently, the well-known MPI standard must be further developed. "The increasing complexity of systems and applications calls for new optimisation approaches," Elis explains. "The idea is to have a common tool to control and analyse the most used functions."

Adapting MPI to new technologies

The scientist finds inspiration for her work above all when she travels, in Corona times rather when she reads. "Travelling means getting to know other cultures and ways of thinking, and that's also important in research." That's why she also appreciates the diverse, international working environment at TUM. Even though she misses her family, friends and her piano in Turkey very much - staying mobile, being on the road, are more important to her at the moment. Her dream job? After her PhD, it will probably be in science: "I want to continue researching," says Elis, "but I don't want to be limited by  corporate policy.

The scientist in the TUM team has already come a good step closer to realizing a better approach for MPI. She focuses her research on the MPI tools and optimizations. Elis regularly attends the meetings of the European and International MPI Forum, where she discusses proposals for changing the MPI standard. To this end, she also engages with former colleagues at the Lawrence Livermore National Laboratory, where she studied the Sierra supercomputer for six months after completing her Master's degree. This has already enabled them to adapt the MPI controlling tool PMPI to new technologies at TUM: QMPI is the name of this possible successor, for which the team has already formulated requirements and use cases and implemented a productive prototype. Now they are improving its design and implementation with MPI tools working group. For Elis, the work is far from over: "With a mindset of research," she says thoughtfully, "new questions constantly arise that you want to solve." (vs)

elis

Persons to watch: Daniëlle Schuman, Informatics, LMU