Domain Decomposition vs. Master-Slave in Apparently Homogeneous Systems

This paper investigates the utilization of the master-slave (MS) paradigm as an alternative to domain decomposition (DD) methods for parallelizing lattice gauge theory (LGT) models within distributed memory environments. The motivations for this investigation are twofold. First, LGT models are inherently difficult to parallelize efficiently with DD methods. Second, DD methods have proven useful for homogeneous environments, but are impractical for heterogeneous and dynamic environments. Besides, many modern supercomputer architectures that look homogeneous (such as multi-core or SMP), are in fact heterogeneous and dynamic environments We highlight this issue by comparing a traditional first-come first-served MS implementation to a simple but yet efficient selective MS scheduling strategy that automatically accounts for system heterogeneity and variability. Our experimental results with the parallelization of our LGT model, reveal that the selective MS implementation achieves good efficiency, but lacks of scalability. In contrast, the DD method is highly scalable, but at the expense of a poor efficiency. These results open up for a hybrid approach, where the MS and the DD methods would be combined for achieving scalable high performance.