NUMA-aware CPU core allocation in cooperating dynamic applications

Building large scientific applications by composing multiple smaller applications is one of current research directions. If the individual component applications are executed together on one compute node, we need to allocate resources to the components. While the operating system is already capable of doing this, it might be possible to get higher efficiency with a more specialized solution. Our goal is to describe opportunities and challenges faced by anyone designing such a system. We look into the specific case where a dynamic runtime system (or multiple such runtime systems) are used by the applications, since we believe the fundamental design of these runtime systems makes them especially suitable for the role. The ideas described in this paper are based on our prior experience in building such a runtime system and our early experiments with cooperating applications. We will focus on CPU core allocation and point out the importance of making non-uniform memory access architectures a prime consideration in such work.