2021-05-31-Person-to-Watch: Amir Raoofy_eng

They are researching future technologies, developing innovative technology or algorithms, advancing supercomputing, artificial intelligence methods or quantum computing: here, in loose succession, we introduce young scientists who are sure to be heard from in the future. Computer scientist and engineer Amir Raoofy is currently doing his doctorate at the Technical University of Munich (TUM) and is exploring how parallel systems and supercomputers process data. He is experimenting on the supercomputers of the Leibniz Supercomputing Centre and on the BEAST test field to optimise and accelerate their performance.

SuperMUC

Racks of SuperMUC-NG of Leibniz Supercomputing Centre, Garching. Foto: V. Hohenegger/LRZ

Orchestrating thousands of compute nodes

High Performance Computing (HPC), says Amir Raoofy, basically works like a choir. For a song, dozens of voices would have to kick in simultaneously, each hitting the right note on its own, always singing to the same beat. "The same effects are at work in parallel HPC systems, where thousands of compute nodes are supposed to work together and execute an algorithm as harmoniously as possible, without delays, in the same beat, without failures," says the doctoral student at the Chair of Computer Architecture and Parallel Systems at the Technical University of Munich (TUM), describing an ideal for which experts in supercomputing are working hard. Just as conductors train voices individually or in groups, engineers orchestrate the work of HPC systems by changing the programming of processors, the communication of different components, the cabling and the programs and applications to be executed, and much more.

This "hard part of IT" is, what gets Amir Raoofy excited. An engineer and avid piano player, he works with sensors that provide data on functions of thousands of computing cores, programs control codes for processors or accelerators using the Messaging Passing Interface (MPI) or High Level Sythesis (HLS), and researches IT technology. He experiments at BEAST, the testbed of the Leibniz Supercomputing Centre (LRZ) and explores LRZ’s supercomputers, which also compute his results: "We are in the data age, I am interested in what measurements or information can be used to best describe or analyze a cyberphysical system in order to better understand its work, intervene when errors occur, and achieve more performance." Using data from gas turbines and from power plants, Amir Raoofy has been investigating how parallel HPC systems handle measured values in time series and how they use patterns to detect failure symptoms since 2018.

Better tuning the technology for supercomputing

It all sounds very specialized, but his doctoral thesis, "Investigation of Parallel Computer Architectures in the Analysis of Large Time Series Data Sets," actually addresses the fundamentals of data processing and how HPC systems work. The results are certainly of interest to power utilities or turbine manufacturers, also in failure analysis in mechanical or automotive engineering, and wherever systems are constantly observed and evaluated, supercomputers themselves. Here, Amir's findings could lead to performance boosts in running applications, perhaps a more efficient use of processors: "This year I've sharpened the thesis, now I'll flesh it out, and if all goes well, I'll be done with it next year."

It'll work out: Amir Raoofy's resume suggests a fair amount of determination, pragmatism and foresight. A consequence of his passion for chess? He broods over the strategy game to relax – a game that was developed in his homeland, Iran, was banned there in the 1980s and is extremely popular again today. Born in 1990 in southern Iran, in Ahwaz, Amir Raoofy studied mechanical engineering at Tehran's Amirkabir University as of 2008, and for his bachelor's thesis he investigated how to control and prevent vibration in vehicle engines. "This involved numerical issues, brought me to HPC and the idea of going abroad for my master's," Raoofy says. "The technology and also the technical level in Iran differ a lot, in some areas the country is at the top of the world, but in computer science, unfortunately, you rarely get access to top systems."

North America, Europe and Germany are where the music plays. Here, master's programs are free of tuition. At TUM, they also combine computer science with engineering, which offered him personal perspectives. In addition, his older sister lives in the Netherlands and an uncle in Berlin. And so, since 2017, Amir has been simulating flows in overland pipelines in Munich for his master's degree using semi-implicit Euler methods. "Germany is a good place to study and do research," he says, referring to the technical and material resources, as well as the independence of science. He stays with HPC, with power generation and also with fault detection; in his PhD, the engineer from Iran brings everything together doing basic research.

Instrumenting Matrix Profiles for Supercomputing

He describes himself as tenacious, crisis-proof, capable of learning. In conversation, Amir is humorous, optimistic, witty. He likes to laugh, sometimes thinks longer about personal and political issues. "Amir is always in a good mood and spreads good cheer," says Dr. Josef Weidendorfer, who heads the Future Computing program at the LRZ and helped develop the BEAST internship, which Amir enriches as a tutor and with his own research questions. Here, students learn about new IT technology and are expected to exploit it in various tasks. Amir can observe how different programming of processors or technical modifications affect the data processing of the systems. Above all, however, the specialist is always successfully testing and tinkering with known algorithms and codes: for example, as part of the research team for the federally funded "Envelope" project, he adapts the HPC code "LULESH" to a specially developed method for error detection and handling. "Amir is always open to technical discussions and interested in technical problems and solutions," says Weidendorfer, praising the collaboration. "When an improvement opportunity is discussed, Amir has it implemented and done shortly after."

They also appreciate his speed and ideas in the TurbO and SensE projects. These also involve the evaluation of large amounts of data, specifically those from power plants and their electricity loads. Amir is discovering ways to make matrix profile algorithms, which search time series for patterns but do not work on HPC systems, sound on supercomputers. The team succeeds in using its (MP)N approach to direct 256,000 of about 311,000 compute nodes, or 86 percent of the SuperMUC-NG's resources. "In our experiment, we performed the fastest and largest multidimensional matrix profile ever computed," Amir reports, not without pride. "We achieved a projected core performance of 1.3 Petaflops." The report of this earns the team the prestigious Hans Meuer Award during the ISC 2020 supercomputing conference. "Amir is extremely conscientious," says Dr. Carsten Trinitis, senior assistant professor, describing his colleague and collaborator. "When I need support, even on short notice, I can always rely on him." Trinitis also enjoys discussing and conversing with Amir: "He tells exciting and unknown stories from his home country and is not only briefed on technical topics."

Publications from different projects, award-worthy results, a wide range of topics for teaching and lectures: This is how careers in research take off. "I'd like to continue doing scientific work," Amir says, "solving numerical and artificial intelligence problems, especially on technology and systems, which then give rise to further research questions in supercomputing." So he will continue to tinker with the ideal in supercomputing: Clocking, orchestrating, and orchestrating technologies, algorithms, and codes so that as many compute nodes as possible in a parallel system work together on an application - just as a choir works on a song. (vs/LRZ)

Amir

Amir Raoofy

More persons to watch you should know:

Mohamad Hayek, LRZ: Computer Science and Datatransfer

Sophia Grundner-Culemann, LMU, cryptography

Bengisu Elis, TUM, supercomputing and programming models

Daniëlle Schuman, LMU, quantum computing