Reactive rebalancing for scientific simulations running on exascale high performance computers

摘要

Exascale computers, the next generation of high performance computers, are expected to process 1 exaflops around 2018. However the processor cores used in these systems are very likely to suffer from unpre- dictable high variability in performance. We built a prototype general- purpose reactive work rebalancer that handles such performance vari- ability with low overhead. We did an experimental validation by devel- oping a reactive rebalancer library in UPC, and using it in a 5-point stencil (heat) simulation. The experiments show that our approach has very limited overhead that compensates for runtime processor speed variations, with or without simulated processor slowdowns.