Visualizing Process Composition and Load Balance in Parallel Coupled Models

Coupled model development presents a set of challenges broadly called the coupling problem; message-passing parallelism complicates matters, resulting in the parallel coupling problem. Performance tuning of parallel coupled systems is complex and performed largely in an ad hoc fashion; from the domain scientist’s perspective the figure of merit is throughput, which is the amount of simulation achieved per unit of wall-clock time. Achieving high throughput for parallel coupled models requires high scalability for each subsystem and compatible combinations of the subsystems’ respective resource allocations (e.g., MPI processes) to minimize idle time surrounding coupling events. Scaling parallel coupled models up to million-way parallelism highlights the need for practical methods for describing and evaluating these systems. I present a a set of complementary tools to analyze and visualize process composition and load balance for coupled models. I state five basic process compositions found in coupled models. Two are the irreducible, well-known sequential and parallel compositions found in common process algebras. I define three new derived process compositions that appear in coupled systems. I define a dynamic load balance hierarchy. I propose a simple graph-based schema for diagramming process composition in coupled models that is capable of expressing dynamic load balance relationships, and I present simple examples illustrating its use. I apply the graphical schema to Version 4 of the Community Climate System Model to estimate the complexity of the process composition and load balance problem for this system.