La WebTV Teratec

S'inscrire à la newsletter >

A télécharger

Forum TERATEC 2016 - David KEYES, King Abdullah University of Science and Technology

Solver software infrastructure for exascale applications
David KEYES, Director, Extreme Computing Research Center, King Abdullah University of Science and Technology

At the heart of the G-8’s International Exascale Software Project (launched in 2009) is the belief that a vast array of applications share a core of tasks that can be abstracted and layered in such a way as to be served by a common software infrastructure, more efficiently than if each community develops their own independently. This philosophy is embedded in the “DNA” of the speaker’s institution, KAUST (also launched in 2009), especially with respect to solver software, represented by the speaker’s Extreme Computing Research Center (ECRC). What is general enough to be leveraged over many applications, however, needs to be tuned to specific hardware. Without co-design of software and hardware, performance of future applications may never exceed today’s few Petaflop/s.  The algorithmic adaptations required to migrate today’s successful “bulk synchronous” open source parallel scientific sof

tware base to the exascale environment include:

  • reducing synchronization scope and frequency,
  • reducing memory traffic per core,
  • exploiting more instruction-uniform concurrency, and
  • relying more on algorithms for fault tolerance than expensive hardware redundancy. For solvers, attention can be confined to a handful of algorithms that possess optimal complexity: lin-log scaling in problem size. Otherwise weak scaling to a billion cores is not practical. 

We illustrate with examples from ongoing ECRC research.

Date d'ajout28/06/2016

Tous droits réservés 2014 © Manufacturing | Mentions légales