Mapping Applications to the Hardware Portably and Transparently

When we consider the grand challenges addressed by distributed systems, we likely imagine large-scale machines and parallel code. Yet, these two pillars of computing – hardware and software – are not enough to ensure high efficiency and reproducible performance. When unaware of the topology of the underlying hardware, even well-designed applications and software libraries can fail to achieve their scientific goals. Affinity – how software maps to and leverages local hardware resources – forms a third pillar critical to computing systems.

Multiple factors motivate an understanding of affinity for HPC and Data Science users. On the software side, applications are increasingly memory-bandwidth limited making locality more important. On the hardware side, today’s computer architectures offer increasingly complex memory and compute topologies, making proper affinity policies crucial to effective software-hardware assignments.

In this talk, I will present mpibind, a memory-driven library to map parallel hybrid applications to the underlying architecture transparently from the point of view of applications. This library provides a simple interface for computational scientists and results in a full mapping of MPI tasks, threads, and GPU kernels to hardware processing units and memory domains. Furthermore, scientists do not have to deal with intricate details of the hardware topology and thus increasing their productivity. Finally, mpibind is portable across computer architectures bridging the gap between performance and ease-of-use on parallel clusters.