In the CRESTA project the software challenge is tackled. In CRESTA the hardware is not ignored. The team understands that exascale machines are hugely complex with complex memory hierarchies and many, many millions of parallel threads but the real challenge is what disruptive innovations are necessary to make codes execute properly on an exascale system. The six codes CRESTA is looking at are weather forecast, fusion modelling etc. and these codes are exactly the sort that have to run on exascale computers.
Mark Parsons went on to tell about the DEEP and DEEP-ER projects. Both projects focus on hardware and software code design whereas CRESTA focuses entirely on software. The exascale challenge is twofold. Researchers not only have to concentrate on the applications’ side but also have to built new and innovative hardware. DEEP and DEEP-ER are dealing with the challenge by bringing together two types of parallel computers: on the one hand the traditional massively parallel computer and on the other hand the accelerated computer, consisting of GPUs or in this case, the Xeon Phi architecture from Intel.
DEEP and DEEP-ER are exploring how to bring these two types of architecture together to build an exascale-ready architecture.
The MONTBLANC project complements DEEP and DEEP-ER. One of the biggest challenges of exascale is obviously the power budget. Today’s technologies would lead to parallel computers that would have 500 megawatt power requirements and we simply can’t manage that, stated Mark Parsons. The challenge is to see how to deal with energy efficiency and this is the focus of the MONTBLANC 2 projects. These projects are taking traditional mobile technologies and deploying these for use as powerful supercomputing components in systems that are built from many of these processors to provide petascale performance and subsequently exascale performance.
MONTBLANC and MONTBLANC 2 have been looking to take the ARM Samsung technology which is used in their mobile devices and connecting them to network hardware and then building massively parallel supercomputers from these components that fit within a very strict energy budget. At the same time, the team has been looking at the whole applications’ and software environment needed for this, from the operating system to the tools and parallel programmes, up to the applications themselves.
The three final projects that Mark Parsons talked about, are three smaller projects, funded after DEEP, MONTBLANC and CRESTA. They are focusing on very specific aspects of exascale. They are EPiGRAM, NUmexas and EXA2CT. These three projects focus much more on the programming model side of the exascale challenge.
EPiGRAM looks at traditional programming models which are widely used today at the petascale such as MPI, OpenMP and the PEGAS programme model and looks at how they can be used at the exascale, so committing the best bits of each of these models and combine them to form a compelling high performance, highly scalable way of programming exascale computers.
NUmexas looks at the industrial applications and the numerical methods required at the exascale. There are many challenges today on the industrial use of supercomputing where it’s already suffering from the inability to scale the thousands of cores. NUmexas looks at how the researchers have to change their numerical methods in order to calculate some things that we really want to calculate at much higher scales, according to Mark Parsons.
The EXA2CT programme is looking at how programming models and software architectures can be brought together in order to tackle some of the biggest challenges we have at the exascale. How do we design codes to make them properly execute at the exascale and so many million of parallel threads?
More information is available at http://exascale-projects.eu .