Horst Simon cites from an Exascale Townhall meeting about the sofwtare findings: Exascale computer architectures necessitate radical changes to the software used to operate them and the science applications. The change is as disruptive as the shift from vector to distributed memory supercomputers 15 years ago. The current message passing coupled with sequential programming languages will be inadequate for architectures based on many-core chips. The present code development, correctness, and performance analysis tools cannot scale up to millions of threads that will run in an Eflop/s computer. With perhaps millions of cores, Eflop/s computers are bound to fail on at least small parts, very often. But the current checkpointing will be inadequate for reaching fault tolerance at the exascale.

The Eflop/s systems will also produce massive amounts of data. Fundamental changes are necessary to manage and extract knowledge from this tsunami of data.

Software challenges for exascale systems are numerous. For instance the creation of development and formal verification tools integrated with exascale programming models is needed to improve the robustness and reliability of the system and the applications. New fault tolerance paradigms will need to be developed and integrated into both existing and new applications. Developers also will have to integrate knowledge discovery into the entire software life-cycle. One just cannot assume that all results needed will be produced and that the results that are produced are correct. Work flow integration, runtime steering and visualisation are areas that also need improvement. Provenance needs to be captured automatically and new effective formats for storing scientific data need to be developed.

Meantime we cannot devote all of our HPC efforts to exascale systems. Though the first supercomputers have crossed the Pflop/s barrier, we still are not at a level were day-to-day sustained Pflop/s performance is delivered to many types of applications. Horst Simon sees no coherent Petascale software plan across different platforms and different agencies (in the USA).

As an illustration, the new supercomputer Sequoia comes with a complete open source software stack that helps in porting applications to this Pflop/s system. The idea is to create a seamless environment that can target the Desktop, clusters and the Pflop/s supercomputer.

The Sequoia supercomputer will be delivered by IBM and installed at the Lawrence Livermore National Laboratory. The initial system, delivered this year will have 0,5 Pflop/s performance. The final installation will reach 20 Pflop/s in 2011.

The USA are spending a lot of money each year on HPC. For the fiscal year 2009, Horst Simon estimates that it amounts to probably about $2B total: High End Computing Infrastructure and Applications get $1,142 M and High End Computing R&D $492 M.