SCREMS: Next Generation Parallel Computing Infrastructure for Numerical Relativity

罗切斯特理工学院数学科学学院计算相对论和引力中心将购买一个计算机集群，用于支持数学科学研究。该设备将用于几个研究项目，特别是包括：发展新的，高效的，极高精度的数值算法，以解决非线性方程的广义相对论磁流体动力学（GRMHD）。 磁流体动力学是一个困难的挑战，由于小规模的特点，在流体材料和磁场的冲击引入数值。这个问题只能有效地解决使用高阶有限差分方法（大于四阶），包括那些设计来发展保守的系统，结合高度可扩展的自适应网格细化（AMR）算法。 在不久的将来，下一代多核处理器有望拥有多达32个独立内核，而拥有超过64个内核的计算节点也将很快出现。 那么，挑战就在于如何在这些众核系统上高效地并行化算法。我们将通过openMP和MPI的结合优化节点间和节点内的通信来实现这一点。建议的集群，最初将有12个节点（在第二年升级到16个），每个节点包含8个核心和16 GB的RAM，与高速，低延迟的InfiniBand网络互连，将模拟这些未来的集群，以便我们可以开发新的算法。该集群将作为未来3年开发数值技术和进行GRMHD模拟的中心工作站。该奖项的部分支持来自物理学，引力物理计划。

The Center for Computational Relativity and Gravitation in the Schoolof Mathematical Sciences at the Rochester Institute of Technology willpurchase a computer cluster which will be dedicated to the support ofresearch in the mathematical sciences. The equipment will be used forseveral research projects including in particular: the development ofnew, highly efficient, extremely high accuracy numerical algorithms tosolve the non-linear equations of general relativisticmagnetohydrodynamics (GRMHD). Magnetohydrodynamics represents adifficult challenge numerically due to the small-scale featuresintroduced by shocks in the fluid material and the magnetic fields.The problem can only be tackled efficiently using higher-order finitedifferencing methods (greater than fourth-order), including thosedesigned to evolve conservative systems, in conjunction withhighly-scalable adaptive-mesh-refinement (AMR) algorithms. The nextgeneration of multi-core processors are expected to have up to 32individual cores in the near future, and computational nodes with morethan 64 cores will soon follow. The challenge, then, is toefficiently parallelize the algorithm on these many-core systems. Wewill do this by optimizing inter and intra-node communication via acombination of openMP and MPI. The proposed cluster, which willinitially have 12 nodes (upgraded to 16 in year 2), each containing 8cores and 16 GB of RAM, interconnected with a high-speed, low-latencyInfiniBand network, will model these future clusters so that we candevelop new algorithms. This cluster will act as the central workhorsefor developing numerical techniques and performing GRMHD simulationsfor the next 3 years.Partial support for this award is from PHY, Gravitational Physics program.