Efficient deadlock-free asynchronous approaches for a distributed-memory sparse direct solver (ICIAM 2015)
Résumé
We describe how to enhance parallelism in an asynchronous distributed-memory environment with limited memory dedicated to communication. In order to maximize asynchronism, we characterize deadlock situations and establish global properties to prevent or avoid them. We also characterize some communication patterns and define a class of broadcast trees ensuring good efficiency for series of successive asynchronous broadcasts. The impact of this work is illustrated on asynchronous sparse multifrontal solvers but has a larger scope.