The challenging problems arising from fast parallel N-body simulations became a driverfor high performance computing. The Barnes-Hut tree code is an example in the class offast summation algorithms, with a complexity of O(N log(N)), instead of O(N2). Themulti disciplinary code Pepc – the ’Pretty Efficient Parallel Coulomb solver’ – is basedon the Hashed-Oct-Tree scheme and is developed at Juelich Supercomputing Centre.The pure bookkeeping overhead of the data-distributed tree construction decreases theperformance and rapidly increases for large scales, as shown for JUGENE, an IBM BlueGene/P architecture. For this reason, novel approaches will be established and applied,minimising integral bottlenecks. An axiomatic and provable optimisation of the parallelorganisation structure, induced by a distributed memory machine, is introduced in detail.Reducing memory footprint and communication alike, the new concept intrinsically guidesto a tight a-priori estimation of parallel data overhead. Moreover, the influence of thelocality-preserving Hilbert-curve on the irregular communication structure, is studied.Accordingly, the new method provides an immense upgrade for the particle number,making Pepc a more versatile tool for simulations in a multi disciplinary context
展开▼
