Position in Near-Memory Computing - Applications: analysis, modelling and run-time managementJointly supervised by Eindhoven University of Technology, Department of Electrical Engineering, Electronic Systems group, IBM RESEARCH GMBH, Switzerland, and ASTRON, The Dutch Institute for Radio Astronomy. The position is a research position that can be used to work towards a PhD or a PDEng degree. A PDEng degree is a professional degree focused on system design and engineering of highly complex system. It prepares for an industrial career. See
https://www.tue.nl/en/education/tue-graduate-school/ for more information about the degree programs.
Project Scope
The computing demands of processor systems are continuously increasing due to emerging application domains, like big-data and deep learning applications. E.g. the next generation radio telescope of ASTRON and others, called SKA (Square Kilometer Array, containing a huge set of telescopes), generates 10 times more traffic then the complete global internet today. In deep learning, many layers of neurons are used with up to billions of coefficients (i.e. weights to be learned), which have to be heavily accessed (for the learning and the classification steps). These applications not only require tremendous processing power, but have also huge data bandwith requirements. This results in memory, performance and energy bottlenecks. This is the area where NeMeCo wants to contribute. NeMeCo researches new ways of computing, which aim at solving these bottlenecks by keeping processing much closer to memory, having e.g. HW accelerators closely attached to memory.
The challenge
Design new generations of HPC systems that can keep up with the ever-growing data volumes, is that current technological trends prevent large amounts of data to be transferred at acceptable power dissipation costs. Reducing expensive data transfers by 'bringing computation closer to the data', also known as near-memory computing, has emerged as a very promising solution to address this scaling issue in HPC systems, in order to realize the Exascale computing systems that are required for handling future Big-data workloads. Near-memory computing, however, is still in its infancy. Many challenges have to be addressed before it can be established as an integral component of HPC systems. Challenges include:
Analysis of applications with big data workloads. This includes data and compute complexity, locality, parallelism, etc.Study architecture solutions and propose new, near memory based, computing architectures.Advanced mapping, algorithmic optimizations and code generation techniques E.g. to increase data locality and optimizing reuse, advanced loop transformations are needed (like tiling, loop interchange, fusion). In addition partitioning and parallelization, like vectorization is needed.Run-time optimizations. At run-time application behavior may change, requiring run-time adaptation of e.g. its mapping, or of architectural settings.Design Space Exploration: this includes modeling of these applications, modeling of proposed computing architecture solutions, and modeling of the mapping of applications on these computing architectures.Build a real system, containing processing hardware and memory, tools, run-time system, and running applications. Typically building FPGA prototypes will first be realized.The project
The successful candidate will work on the project NeMeCo - Scaling Big-Data Processing into the Next Decade. NeMeCo is an ambitious Marie Sk³odowska-Curie European Industrial Doctorates (EID) Innovative Training Network (ITN) programme, which addresses several of the above challenges. In particular, NeMeCo is an interdisciplinary training and research project aimed at developing power-efficient HPC systems for Big-data processing based on the exploitation of near-memory computing capabilities, and, in this way, making the world a better place by enabling important innovations in, for example, healthcare, energy consumption, traffic congestion and safety.
NeMeCo involves a three-partner network comprising the Eindhoven University of Technology in the Netherlands, IBM Research GmbH in Switzerland, and the Netherlands Institute for Radio Astronomy, ASTRON. These partners provide a unique combination of academic and industrial expertise of Big-data applications, compilers, memory and processor technology, and high-performance computer architecture.