|Name:||Simulation Laboratories at JSC: Community-Oriented Software Support & Development for HPC|
|Time:||Tuesday, June 19, 2012
4:00 PM - 5:00 PM
CCH - Congress Center Hamburg
|Speakers:||Paul Gibbon, Jülich Supercomputing Center|
|Abstract:||Over the past two years, a road map has emerged which, despite stagnating processor clock speeds, foresees the advent of exascale machines capable of 1018 floating-point operations per second by 2020. This can only be achieved by a combination of increased concurrency, drastic reductions in power consumption, and a complete rethink of processor memory access. These challenges have prompted a host of exascale initiatives and consortia aimed at addressing the major hardware design issues.
For the computational scientist, these emerging architectures once again necessitate a major paradigm shift, from pure distributed memory algorithms (MPI) to more complex program structures featuring concepts such as hybrid communication layouts, asynchronous tasking and non-blocking I/O. Just a decade ago it was still possible to develop complex theoretical models without too much regard for algorithmic or numerical elegance: the raw number-crunching power of available processors would eventually compensate for sub-optimal code design. To fully exploit contemporary supercomputers, simulation scientists are compelled to take scalability into account when developing new models, or when reengineering established codes to new machines.
This poses significant problems for small research groups making up the vast majority of supercomputer facility users, who generally do not have the resources or expertise for serious application ‘Peta/Exascaling’. Many important community codes are typically able to exploit only 5-10% of the available computing power because of their poor parallelism, and there is a tendency for these applications to get ‘left behind’ as HPC hardware evolves further. On the other hand it is vital for the sustainability and mandate of leadership computing centres that their scientific user bases remain as large and diverse as possible.
To help meet this challenge, JSC has established several ‘Simulation Laboratories' providing high-level algorithmic expertise in a number of strategic disciplines. These teams provide a bridge between the user groups and current/future production systems. The structure and status of the JSC simulation labs will be reviewed, highlighting recent research as well as a recently concluded advanced support initiative targeted at selected communities.