ASTRON & IBM Center for Exascale Technology

Algorithms & Machines

Until a few years ago, the design of computer systems was relatively simple: there was a high speed CPU plus memory and the majority of the optimization was inside the CPU chip. As scaling has basically stopped and today, for designing computer systems, a wealth of parameters and boundary conditions need to be observed: power envelope, performance, hardware cost, workload, CPU, memory size, accelerator technology, boards, cooling, system etc. This has propelled the complexity of the design far beyond what a reasonable team of people can keep “in their head” and reason about. IBM Zurich has embarked on developing a methodology, based on mathematical ground principles which addresses the fundamental equations which govern the system and uses mathematical optimization to give a couple of directions system design should address. The methodology will allow to reason on system optimizations and parameter space. This is as key for IBM’s own future system designs as for large research infrastructures like the SKA.

As a first step, we will perform a retrospective investigation using the LOFAR (and possibly the South-African Meerkat telescope) design parameters, comparing the actual system design of this SKA PathFinder and what the tool (currently in development) would suggest. ASTRON has already investigated their algorithm decomposition with the result that e.g. the amount and access to memory seems to be the most influential parameter. However, the detailed decomposition of the problem and mapping on available platforms, including accelerators for specific functions, is beyond the current methods and tools. For ASTRON, being exposed to such a tool in an early phase and having insight in its development and operation would be a major differentiator in skills they can bring to bear on the future design choices for SKA. Determining and establishing of the SKA baseline design with realistic parameters will be part of the project. We will kick this off immediately with an ASTRON/IBM workshop involving the key architects of LOFAR and SKA experts from Cambridge University.

This fundamental architecture project will also be an umbrella for the remaining 6 workstreams and help with making the proper architecture decisions when appropriate.