Q: What should be done about the applications that won’t be able to exploit thousands of GPUs together in a single simulation, for example because of algorithm limitations or legacy coding issues. Will the community be supported to develop and implement new algorithms for those codes? Or will it be acceptable for those codes to use large numbers of nodes but leave the GPUs idle? Blue Waters had a significant planning and preparation effort with the applications development community – what are the plans in this respect for Titan?
Submitted by Andrew Jones, NAG
Q: It is not the thousands of processors that should concern us, but the millions of cores that are just around the corner, and the billions of threads that will be required to exploit these systems efficiently. I believe that the industry needs radical new programming paradigms for Petascale and Exascale systems, and that the paradigms we use today won’t scale – and if they did there is no realistic provision for resilience. The issues that must be addressed are massive concurrency, program hierarchy, heterogeneity, program portability and application resilience. What programming paradigm will address all of these issues?
Submitted by John Barr, 451 Group
For related stories, visit The Exascale Report Archives.
