PRACE to unveil six Petaflop prototypes
Over the next few weeks, PRACE, Partnership for Advanced Computing in Europe, will announce its choice of six Petaflop-class prototype systems and the sites that will host them. The set of six prototypes are expected to cover a wide variety of the most promising next-generation HPC architectures. In preparation for the deployment of these systems, the project has also conducted two major studies. The first study has identified a suite of applications that represent a cross section of application areas across Europe, which will be used in benchmarking and optimisation studies on the selected Petaflop/s systems. The second major study was a European-wide survey of HPC training and education needs, which will be used to steer user training for future Petascale systems.
The set of prototype systems is expected to cover a wide variety of advanced computing architectures that are expected to scale to Petaflop/s performance in the next one to two years. In preparation, these prototypes will be assessed at many levels including performance, scalability, total cost of ownership and energy consumption. It is expected that the prototype systems will be deployed at different PRACE partner sites by the end of 2008. The chosen hardware, and the sites where the hardware will be housed, will be announced in the near future.
Meanwhile, PRACE has conducted two major European-wide end user studies. The first study had the goal of identifying which application areas and codes were representative of European HPC usage and potential candidates for future Petaflop/s systems. The survey data was used to produce an overall utilisation matrix characterised by scientific area and algorithm. The result of this is a core list of nine applications, which are representative of the overall utilisation matrix, plus a set of 10 possible extensions. These applications will be ported, benchmarked and optimised on the prototype machines.
The second study was an evaluation of training and education needs of prospective PRACE users. Top-tier users at each PRACE site were invited to participate in an online survey that examined their proficiency and training requirements in a broad range of HPC areas including MPI (Message Passing Interface), OpenMP (Open Multiprocessing), hybrid MPI-OpenMP programming, PGAS (Partitioned Global Address Space) languages, parallel I/O, visualisation, parallel debugging, performance optimisation, multi-core programming, numerical libraries and distributed computing tools. The survey highlighted strengths and weaknesses in current HPC training and revealed the need for centralised HPC training repositories and HPC technology dissemination channels. The results of the survey will be used to steer future training and education programmes within the PRACE HPC research infrastructure.