CRESTA project to figure out how to keep more than 100 million threads busy

18 Jun 2013 Leipzig - In the first BoF session on European exascale research at the ISC'13 event, Erwin Laure from KTH in Stockholm, Sweden, presented the EC-funded CRESTA project. CRESTA fouses on the software challenge in the search for exascale computing. Hardware solutions are being developed to deliver exaflop performance but developers face unprecendented levels of heterogeneous parallelism. CRESTA is trying to put a finger on what the scientific challenges are and on what new modelling and algorithmic methods are needed. Erwin Laure is well aware of the complexity and asked himself how to keep more than 100 million threads busy. His talk made clear that hardware and software are inextricably linked.

The CRESTA project basically is about software for exascale computing. CRESTA stands for Collaborative Research into Exascale Systemware, Tools and Applications. The project team is developing techniques and solutions.

Among the partners are leading computing centres, vendors, and application owners and specialists. This is the focus of the work in CRESTA. The application owners and specialists are enabling applications towards exascale computing.

Erwin Laure told the audience that the road to exascale is complex. You do not get there just by simply building hardware. The vendors understand this. In the USA, they have already stopped talking about exascale and they are focusing on extreme computing.

Erwin Laure also said that the technologies we need are becoming clearer. Together with CRESTA, the DEEP and Mont-Blanc projects are also funded by the EC to develop these technologies.

The speaker also asked himself how to live with Gene Amdahl. This is difficult and requires to observe a few rules. The basics are to minimize the sequential part and to maximixe the parallel part. The challenge is to make "p" as parallel as possible.

The developers are also fighting with the problem to minimize communication between sequential and parallel and to try within the parallel to improve the application software and algorithms, and the systemware as well.

Erwin Laure is convinced that Ensemble Computing has its place at the exaxscale level. The developers have to look at moecular dynamics or numerical weather forecasting for proof. For some applications however they need to consider how to make them scale to more than 100 million parallel threads. For these applications they need to rethink how they can model and simulate their physical problems. In many cases they also need to rethink how they can express differential calculus on these compting devices. Erwin Laure thinks that far too much exascale research is focusing on incremental improvements to algortithms.

He summed up the CRESTA principles. CRESTA is a two strand project: it is building and exploring appropriate systemware for exascale platforms while also enabling a set of key co-design applications for exascale.

The co-design is at the heart of the project. The developers try to provide guidance and feedback to the systemware development process. They have to integrate this development and benefit from it.

The task is to employ both incremental and disruptive solutions. Exascale computing requires both approaches. The incremental approach is realized through optimisations, performance modelling and co-design application feedback. The developers are looking to achieve maximum performance at exascale level and understand the limitations. The disruptive approach works with co-design applications and has to consider alternative algorithms.

Within CRESTA six appiications are being developed including weather forecast, Elmfire, GROMACE, OpenFOAM, and HemeLB.

More information is available at the CRESTA project website.
Leslie Versweyveld