Keywords: high performance computing, architectures, programming, energy.

In recent years high performance computing has seen a dramatic change. First and foremost we have seen a change in architecture. Back in 2002 the basic architectural concept was the cluster. The building blocks for high performance computer architectures were standard components. Big clusters were expected to dominate the market for a long time to come. The promise for the future was a single, consistent and standardized architecture that would span a range from desktop system to high end computers.

Looking at the landscape and planning in 2008 the picture has changed completely. The TOP500 list of fastest systems in the world is dominated by systems like BlueGene with a completely different architectural approach. At the same time we see future planning in large scale projects that aim at creating so-called hybrid architectures. The most important projects are Cascade, RoadRunner and the Japanese Next Generation Supercomputer Project. Rather than providing the community with a single platform, hardware vendors offer a basket of solutions that follow entirely different approaches. As a consequence the community has to adapt codes and methods to an ever more volatile hardware base. It is only a side effect that while playing around with new architectural concepts vendors have entirely ignored the power consumption issue. Costs for power supply have at least grown by a factor of three over the last 5 years. This is not to be counted as the least problem for high performance computing simulation. The operational costs of hardware have grown to become as high as the initial investment costs - for some medium sized centres prohibitively high costs.

While the computer science community is struggling with these problems simulation experts are faced with a different challenge. The level of performance of high performance computers is so high that traditional simulations can be done easily. At least there is no more performance barrier for some of the more common problems. As single phenomena (like CFD, crash, ...) can be simulated on a reasonably large system the trend is towards coupled application. In such applications users have to deal will multi-scale problems which make further programming difficult.

purchase the full-text of this chapter (price £20)

go to the previous chapter
go to the next chapter
return to the table of contents
return to the book description
purchase this book (price £98 +P&P)