Supercomputers, which were a long time ago built on technology developed from scratch, are nowadays built from commodity components. On one side, this means that designers of such systems have to closely monitor the evolution of mass market developments. On the other side, supercomputing becomes a driving boost for technology and systems for the future, providing requirements for the performance and design. As fundamental limits in the single processor per chip, in terms of performance/power ratio, are already on the table, multicore chips and massive parallelism have become the trend to achieve the required performance levels. A hierarchical structure, both in hardware and software, is the unavoidable approach build future supercomputing systems. The talk will first address how these systems have been built in the past and how we envisage their design in the near future.
The gap between peak and real performance for current systems will become worst if designers don't adopt a vertical approach, from processor to node and system design (including interconnect), parallel programming models, dynamic resource management to improve load balancing, tools for performance analysis, prediction and optimization, and new numerical methods, algorithms and applications. Research and proposals in this direction will be presented during the talk in the framework of the future 10/100 Petaflops architecture for the Marenostrum site, the Barcelona Supercomputing Center.