WG211/M15Scholz

This talk presents a new take on Distributed Shared Memory (DSM). It is based on the idea that compiler generated parallel codes can guarantee certain invariants which reduce the overhead of DSM significantly. Based on a prototypical implementation named SHRAY, we present some initial performance results. They suggest that a performance similar to that of hand-optimised PGAS implementations can be achieved while offering a much more flexible programming model.