DNS at extreme scales: two new approaches towards reducing cost

The wide range of spatial and temporal scales in realistic turbulent flows makes it extremely difficult to perform simulations that can capture the complete flow physics accurately.

Because of the extreme computational cost of these simulations, they are typically done on the most powerful massively parallel supercomputers available. This cost stems from both the exceedingly large number of degrees of freedom that need to be resolved as well as the cost of moving data between processing elements-necessary due to the tightly coupled nature of the equations of motion.

In this talk we present two new approaches to reduce the enormous cost associated with (i) the massive number of degrees of freedom, and (ii) the overhead due to data transfers at extreme scales. For (i), we show recent results from Selected-Eddy Simulations (SES) which solves Navier-Stokes (NS) dynamics for a subset of modes over the entire spectrum of

scales and evolve the rest according to trivial (computationally cheap) dynamics. We show that for turbulent mixing, for example, less than 1% of NS modes are needed to capture accurately the physics at all scales. For (ii), we show the recent development of asynchrony-tolerant (AT) schemes which reduce the number of communications while still preserving physical accuracy. We show how both approaches can be combined and reduce the energy requirements for accurate and realistic simulations At massive scales.