The impact of global communication latency at extreme scales on Krylov methods

Krylov Subspace Methods (KSMs) are popular numerical tools for solving large linear systems of equations. We consider their role in solving sparse systems on future massively parallel distributed memory machines, by estimating future performance of their constituent operations. To this end we construct a model that is simple, but which takes topology and network acceleration into account as they are important considerations. We show that, as the number of nodes of a parallel machine increases to very large numbers, the increasing latency cost of reductions may well become a problematic bottleneck for traditional formulations of these methods. Finally, we discuss how pipelined KSMs can be used to tackle the potential problem, and appropriate pipeline depths