Trade-offs in Context Identifier Allocation in MPI

The notion of communicators is one of the most central concepts in the Message Passing Interface (MPI), allowing the library developers to contextualize their message exchanges, and scope different algorithms to well-defined groups of processes, and allowing the MPI implementation to specialize its matching and deliver messages in the right context. Due to its collective nature, the communicator identifier allocation is directly linked with the performance of the collective communication implementation on the execution platform. We propose a scalable algorithm for communicator identifier allocation, that minimize the amount of data involved in the operation. This allows for faster context allocation implementations, that are capable of taking advantage of the hardware accelerated collective where available. Additionally, we present 3 communicator storage strategies, and their implementations in the context of Open MPI. We explore the performance of these new algorithms, compare them with algorithms available in other MPI implementations, and discuss their trade-offs.