A Case Study on Multi-Component Multi-Cluster Interaction with an AMR Solver

Abstract—The present paper studies an assembly of parallel applications—a controller, a dynamic adaptive Cartesian mesh simulation code for hyperbolic partial differential equations, and a multi-device visualisation server—where the codes interact via procedure calls formalised by an interface definition language. These calls are augmented by target device identifiers. Answering nodes hence are efficiently able to route their data flows through the multi-cluster environment. We study a strategy where answers from multiple nodes are merged dynamically along a virtual tree either on the sender or receiver side. Dynamic merging allows the data flow to adjust to the communication characteristics of the codes, their load balancing, and the hardware topology. The communication between simulation and postprocessing compo-nents is handled transparently through well-defined interfaces that guide a code generator. The generated communication code, the data scattering and the data reduction are hidden from the applications and its programmers. The assembly realises concepts of the common component architecture. Results from our case study show reasonable usage of available communication facilities. Hence the present work describes a mature architectural pattern for interactive parallel computing and non-holistic high performance software assemblies. Keywords—On-demand data exploration, blackbox AMR, cou-pling, near-data processing, parallel data exchange, parallel com-ponents I