Making Exaflops Work

Sometime in the next decade, probably around 2015, a significant change in HPC systems design and operation will occur setting a new trajectory that will lead to Exaflops computing by the end of this decade. The ground work for this latest revolution is already being laid but its impact will not be applied generally until conventional practices are exhausted. Interestingly, software is likely to precede hardware in these advances reversing usual trends where software has almost always had to play catch up. The major change will move from static to dynamic resource and task management, from message passing to message driven computation, from distributed memory to global address spaces, from global barriers to lightweight futures synchronization, and from compiler to runtime determined scheduling. Heterogeneous systems of up to a billion cores of different types will provide fault tolerance and power management automatically. Even today some applications are limited in their scalability due to conventional programming techniques. New software strategies are already beginning to be explored and these incipient techniques will drive system change. This talk will describe the likely new model of computation that will replace the prevailing message passing model and using recent results from the DARPA studies and other sources will describe likely systems to lead HPC by the end of this decade.