Parallelizing Lattice Gauge Theory Models on Commodity Clusters

This paper addresses fundamental parallel computing issues for efficiently parallelizing 3D Lattice Gauge Theory models (LGT) on distributed memory systems. The long-range application stencil of LGT models put together with the impossibility of updating neighboring lattice sites simultaneously greatly complicates the parallelizing of such simulations. Our algorithms decompose the domain among the processors, and settle a staggered execution with the help of virtual tokens that circulate among the processors, allowing the token holders to update their boundaries. Experimental results show that these algorithms are scalable, and that simple communication trajectories prevail over low surface-to-volume ratios. Rigorous theoretical results are provided under the LogGP model to demonstrate the superiority of our approach over other methods found in the literature