Technical Perspective: Unexpected Connections

International audienceTo improve program performance on today's clusters, clouds and multicorecomputers requires the developer to pay serious attention toscalability. When a thread updates some shared datum, and another thread wants to reador write the most recent version of that datum (they \emph{conflict}),they must synchronise, which constitutes a sequential bottleneck.This is a general result that does not depend on any particularimplementation, even with efficient hardware support for cachecoherence, as explained in the paper.Here comes the paper's main insight.If two concurrent procedure calls \emph{commute} with each other (i.e.,executing them in either order is equivalent), this means that neitherone depends on the result of the other.Therefore, \emph{there is no inherent reason why these calls should conflict}; and, hence, \emph{it is possible to implement them in a way that scales well}